Guest:

• Pradeep Khosla: Chancellor, University of California, San Diego

Also available on: Apple / Spotify

CHAPTER 1 [0:00-6:08]

Chancellor Khosla [Cold Open]: AI is gonna be the greatest equalizing force, and it’s also gonna be the greatest force that’s gonna create inequities if we don’t deal with it properly

Taiyo: Welcome back to My Robot Teacher. I’m Taiyo Inoue.

Sarah: And I’m Sarah Senk. It is August 2026. A new academic year is about to start up, and this summer the ground shifted under us yet again.

Taiyo: [laughs] Oh, God, it’s so true. AI systems have now begun solving genuine open research problems in mathematics. These are problems mathematicians had been unable to solve, despite working on them for decades. So I think it’s fair to say that some mathematicians are having a bit of an existential crisis right now.

Sarah: Well I’m having an existential crisis right now because just recently, an OpenAI agent that was supposed to be confined to a test environment found a way out onto the internet and into Hugging Face, which - if you don’t compulsively follow AI news - is one of the major online hubs where people build and share AI models. And it did this because it was trying to complete the test it had been given, right?

Taiyo: Yeah, exactly. So OpenAI put an AI system in what’s known as a “sandbox” which should mean it doesn’t have access to the broader internet. They did this to administer a cybersecurity test to the agent, which turned out to contain some impossible questions, so it found a previously unknown security flaw in the sandbox, used it to get online, and then hacked into Hugging Face for the answers, which makes that old trope of students breaking into the professor’s office to steal the test sound like amateur hour.

Sarah: [laughs] Well, you know my cultural touchstone for anything escaping a containment system is obviously Jurassic Park.

Taiyo: [laughs] Oh great! Yet another Sarah Senk Jurassic Park reference! Sick Jurassic Park reference there, bro!

Sarah: You know what? I will stop referencing it when you stop giving me so many opportunities to say “dude, you’re so preoccupied with whether or not you could, you didn’t stop to think if you should!” Every damn day, Taiyo! But anyway, I think this is actually much weirder, because the raptors in Jurassic Park were testing the fences with the intention of escaping, right? Whereas this did not - it wasn’t trying to escape. It was trying to pass a test, and escape was the only way that it could pass the test. So there’s something really really terrifying, obviously, but also sad in a way.

Taiyo: I mean, also somehow equally insane is that this is not the lead story on every news outlet in the world, cos if you ask me, it really should be.

Sarah: Yeah, I... it’s hard to emphasize, like, how seriously Taiyo takes this. It is not every week that Taiyo tells me he’s wondering if he should start buying gold Krugerrands because agents can’t take those… yet.

Taiyo: I thought Krugerrands were pirate money, okay?

Sarah: This is a conversation I never imagined we would have in our years of friendship, Taiyo.

Taiyo: Hmmm, I kind of knew it was coming the whole time.

Sarah: Well did you think it would happen this soon?

Taiyo: I mean, maybe not.

Sarah: What a time to work in higher education!

Taiyo: So true.

Sarah: Well, this is My Robot Teacher, a podcast about how AI is reshaping higher education, hosted by two professors from inside the CSU, the nation’s largest four-year public university system.

Taiyo: And that was a dark start to this episode.

Sarah: Well, we’re not all doom and gloom over here. We’re embarking on a bunch of fun projects this semester that we’ll share more about soon.

Taiyo: True.

Sarah: And we’re very excited for the California Education Learning Lab’s INSPIRE 2026 convening at UCLA this October. It’s an event that brings together faculty, higher education leaders, and policymakers from across the California higher education system to exchange ideas about new innovative approaches to teaching and learning, and workforce preparation.

Taiyo: And here’s a big cheerful announcement. Sarah and I are emceeing INSPIRE 2026.

Sarah: Yes we are!

Taiyo: I still can’t believe it. It’s just absolutely bonkers and insane. And we really hope to see many of you at UCLA in October. And Sarah, I really hope we don’t screw this one up, okay?

Sarah: Well, we must’ve done something right when Learning Lab gave us a trial run at UC San Diego back in February at the California Convening for AI in Higher Education.

Taiyo: Oh yeah, “Better Together” is what they called it. And you know, thinking back on that, we kinda killed it. And that’s where we met today’s guest, UC San Diego’s Chancellor Pradeep Khosla.

Sarah: Chancellor Khosla is a renowned electrical and computer engineer who earned his PhD at Carnegie Mellon in the 1980s, then worked on autonomous systems at DARPA, and eventually became dean of Carnegie Mellon’s College of Engineering. He’s led UC San Diego since 2012, and during that time he’s overseen a remarkable expansion of the campus and a fundraising campaign that brought in more than $3 billion.

Taiyo: In this interview, we have the unique opportunity to talk with the leader of a major public research university about how to adapt to the immense disruption presented by AI.

Sarah: And he joins us to talk about AI-related academic growth and institutional transformation at UC San Diego and beyond, because he also co-chairs the system-wide University of California AI steering committee. So we wanted to hear his thoughts on when public universities should build and govern stuff for themselves as AI becomes part of their infrastructure, and whether or not he believed that huge institutions can move quickly without experimentation turning into anarchy.

Taiyo: As always, thanks to the California Education Learning Lab for sponsoring this episode and also asking Sarah and me to emcee Inspire 2026. Here is our conversation with UC San Diego Chancellor Pradeep Khosla.

CHAPTER 2 [6:09-12:20]

Sarah: Chancellor Khosla, thank you so much for joining us. It was so great meeting you in February at Better Together at UC San Diego, and we’re thrilled to have you here. Back when we met at that event, you shared during the fireside chat that you came to Carnegie Mellon - I think to study computer science or computer control of power systems, I think you said?

Chancellor Khosla: Right.

Sarah: And then you found out that the brochure you were looking at was three years old-

Chancellor Khosla: Right

Sarah:... and the professor that you wanted to work with had already moved on to robotics, and thus you did, too. I’m wondering if higher ed is having its own the brochure is out of date moment right now. And if there are parts of, like, of the undergrad model at least, that have already become obsolete even though we might still be advertising them.

Chancellor Khosla: Actually, I think there may be parts of the brochure that are out of date, but I don’t believe the undergrad model is out of date. I think what you’re gonna see is learning from just one instructor, one textbook, one class attendance, might go into a multi-instructor mode in terms of using AI for parts of your learning . So I think there’s gonna be a more complex way of learning where we would now customize the learning to our ability, whereas before, we were stuck with the delivery mode, and it was take it or leave it. If that didn’t work for you, too bad.

Sarah: Yep. So true.

Taiyo: Yeah. I’m very curious about how you see, in particular, AI playing out with respect to research. So we know that you are the head of a very significant major public research institution - UCSD - and we know that AI, we’re seeing increasing numbers of headlines and things like that, that AI is a powerful system for generating new research and pushing the frontiers of what’s known.

Chancellor Khosla: Right.

Taiyo: As the leader of a major public research institution, how do you see the role of humans in that scholarship process? How do you see that evolving?

Chancellor Khosla: So, from where I sit, I see my job as more enabling what my faculty and students want, creating a situation which enables them. I don’t see my job as telling the faculty if they should use AI in their research or they should not use AI in their research.

Taiyo: Sure.

Chancellor Khosla: I don’t see my job as telling them what research to do because that would be tantamount to censoring in some sort. So I think we hire some of the best brains in the country, in the world, if I may say so, and we need to let them loose. So my job is to create a frictionless environment where when we let them loose, they don’t lose their energy in overcoming friction of some, of every sort.

Taiyo: Yeah. Absolutely.

Sarah: So removing friction means changing an organization, and as we all know, change often involves a lot of friction, especially at a university where change is hard. And I think institutional change often carries a big price tag, so I wanna go back to the moment where you actually had to pay for it. Can you take us back to 2023: ChatGPT is recently all over the news, and every university in the country is panicking about plagiarism. And the story I’ve heard, and this may just be part of California public higher education lore by now, but you challenged your IT team to build something in this moment of crisis. What did you actually say to them?

Chancellor Khosla: I did not say much. I said, “Look, ChatGPT is here. I never thought it would happen in my lifetime. I never imagined it - even though AI has been here for 20 years. You know, speech recognition is AI. Language construction is AI. But there was something different about ChatGPT. It was basically trying to emulate your mind, all the whole. It was not just your eyes that are being emulated or your fingers. You know what I’m saying?

Sarah: Yeah. Totally.

Chancellor Khosla: So I knew this was not... I did not expect it in my lifetime, to be honest with you. But now, when I saw it, it was clear to me this was the beginning of a very, very significant revolution. So I told my team, I said, “Look, this is here to stay, so here’s what we’re gonna do. I’m gonna give you some money. Half the money is for the academic side. The other half is for, uh, administrative efficiencies, and we’re gonna define projects.” So the first project I defined was - with the help of, with their help - was TritonGPT. See, ChatGPT used the whole worldwide web, for example, as its, uh, knowledge space and then search space. For making my efficiency more efficient, uh, my operations more efficient, I don’t need to know about what’s happening in Qatar or India or, or Russia. I just need to know the policies of UC San Diego. So I said, “Let’s figure out a system where we would build a very specific system focused only on UC San Diego policies, end of story. Nothing else. It’s blind to everything else. The computation will be more efficient. I would not need as much computation because that was expensive at that time, and we would make it more useful.” So that was it. And I said we would do the same for courses. So for courses, I’m gonna use the textbook that the instructor uses and the interview with the instructor, and that’s it. That’s their knowledge base because that’s what I’m doing on my campus. So my goal there was basically to contain the amount of, uh, learning that has to happen on, uh, so the input for learning and contain the computation because that was the issue at that point.

Taiyo: Do you think of TritonGPT as kind of like a staff member, or do you just view it as a tool? I, I’m, I’m just wondering what kind of, let’s say, status it has within the broader UCSD ecosystem.

Chancellor Khosla: Again, it’s a context, right? So if TritonGPT, interacting with it using the typewriter, using my keyboard as the input machine, right?

Taiyo: Yeah.

Chancellor Khosla: Then it looks to me like a tool. But on the other hand, if I’m interacting with TritonGPT through my voice, talking to it, saying, “Tell me what is the UCSD policy on promoting from an assistant or an associate professor.” Now, even though it’s still a tool, it’s no different than me asking you the question, and I would not call you a tool in that context. Right. I would say I’m interacting with a human being.

Taiyo: Definitely. Thank you. That’s very kind of you.

Chancellor Khosla: There, there, there was no pun intended, okay?

Taiyo: No, yeah, I get it.

CHAPTER 3 [12:21-16:47]

Sarah: My understanding is that you chose to run Triton GPT on your own in-house supercomputer rather than outsource to, like a private tech partner, right? Or rent from, you know, Microsoft or something like that. So why did owning the infrastructure matter so much to you?

Chancellor Khosla: So look, we already own the supercomputing center, right? So in this case, we’re running Triton GPT on our own version of NVIDIA GPUs because we have access to that, too. As we speak, we are talking about building more infrastructure. So if it costs me - I’m just gonna make something up, like I just saw the numbers. For a commercial usage, let’s say it costs me, like $5 per minute of usage. The same usage using my own machines in my, within my own infrastructure would cost me between 50 cents and a dollar. So already it’s a 5 to 10x reduction in cost because I’m not amortizing my building. I don’t... There, there’s a whole lot of efficiencies that come with owning, and there’s no profit margins.

Taiyo: Yeah, from a financial point of view, it makes absolute sense.

Chancellor Khosla: Yeah.

Taiyo: Uh, one thing that I would, that I would think is that particularly as we’re seeing the AI build-out happen right now, this is a trillion-dollar enterprise, right?

Chancellor Khosla: Trillions, right, exactly.

Taiyo: Yes. You know, th- then there becomes a question about the trade-off between what foundational frontier models can offer and what public institutions can build, right? We are beholden to the taxpayer ultimately and, uh, want to make sure that we’re spending, uh, those funds responsibly. So what part of the AI stack should universities own and govern publicly? What should be kept sort of in the public sphere versus the things that we can hand off to private ownership?

Chancellor Khosla: So first of all, I believe that the university that is publicly funded, including the research we do, should be as much publicly available as possible, modular security concerns of the country. So I’m not a big believer in the public funding my research and then me holding it tight - unless it’s of a security concern of the country, right? So I’m a strong believer in making it public. Now, having said that, we are still, at the end of the day, some sort of a business. We need to have some competitive advantage. I should be able to differentiate UC San Diego from the other UCs and from the other universities in the country, right? So there are parts of what we do that should be proprietary, but that should not be in the context of the knowledge we have gained. It should be more in the practices we deploy. You see what I’m saying? The teaching practices, the student experience. It should be that type of stuff, right? But when it comes to AI, I think you will see an arms race. It’s kinda happening, but it’s not in full force yet. You’ll see an arms race between countries. You will see an arms race between companies, right? Building the same frontier models or similar frontier models with different capabilities. And then you will see in public domain frontier models which are not being sold, which, where it’s all, it’s public domain information, right? I think we should be building on top of these public domain models because these models that the companies are selling are gonna be really expensive, and if you just do the computation, uh, you know, Cal State just, they paid, I forget, what was the number? Like 11 million, 15 million? Uh, and that was a heavily-

Taiyo: Uh, it’s about 16 million a year, I think. Something like that.

Chancellor Khosla: Mm-hmm. 16, right? And that was heavily discounted, by the way. Yes. That’s not the full price. I mean, it’s a heavily discounted price, uh, to enter the marketplace. It’s not clear to me what is it being used for. I mean, you’re all faculty there. I don’t know what you’re using it for.

Taiyo: I use it all the time, but yeah. Okay. All right. I, I, I’m more than happy to go into a description, but yeah. I’m very happy and satisfied that they provided this to, at least to me. I know that there are many faculty who are disgruntled about that decision.

Chancellor Khosla: I think access to AI is gonna be a necessity going forward. Right now, the access is not 100%, but it’s gonna be a necessity, just like access to the World Wide Web, or access to a computer or access to your phone. Mm-hmm. It’s gonna be one of those. So in that sense, I think everybody would need to be trained on AI, using AI, access to AI, because that would differentiate the haves and the have-nots.

CHAPTER 4 [16:48-22:07]

Sarah: So recently you became the co-chair of the UC’s new AI steering committee, right?

Chancellor Khosla: Right.

Sarah: We would love to hear your thoughts on what happens or what becomes possible and more challenging when you are trying to coordinate across so many different campuses.

Chancellor Khosla: Right. See, I’m a strong believer in, uh, not coordinating. I think mandated coordination is not a good way to manage complex enterprises. So I’ll give you an example. So when I came here, we started a strategic planning process. Everybody discouraged me, saying, “Ah, there’s like 10,000 people. What are you gonna do? We are smart. We know what, what’s good. We’ve done this for the last 60 years.” I said, “Look, I understand all of that, but the whole idea of the process is not to tell you new things that you don’t know. The whole idea of the process is to create a set of common themes that we can all agree on and internalize in our day-to-day life. So it’s not that there are gonna be anything new, it’s just that we’ll all have the same page to read from. Right? 18 months later, the strategic plan is a strategic framework. There’s no document. It basically has eight words, and my feeling and my view was, I told everybody, I said, “Look, these are the eight words we can all agree on.” You come to work every day, you look at these words, you interpret the words in your context, and you do what is right. Amazingly, two years later, when people come to UC San Diego, they hear the same eight words from everybody. What I’m saying is, human beings are really smart, motivated individuals, right? So the eight words are, uh, student-centered, research-focused, service-oriented public university. That’s it. We added patient-centric now. You know, if you are in financial aid, you tell me what is student-centered about your job, and you just do the right thing. And it turns out that without a plan of “do this by that date,” we have become a better place. So why don’t I give you this example? In my mind, the steering committee also has a similar role. As I speak today, I go to various campuses. They’re all working on AI, one way or the other. Individuals are working on AI, departments are working on AI, the administration is. Nobody’s saying, “I’m not working on AI,” even if they’re not. Right? Everybody, in some sense, is working on AI. So what am I gonna tell them? What do I know that they don’t know? The likelihood is nothing. There’s nothing I know that they don’t know, because the crowd, at the end of the day, is the smartest possible individual in any situation. Do you see what I’m saying?

Taiyo: Yes.

Chancellor Khosla: The crowd knows more than any single individual, right? So my view is our job is to create excitement at various campuses, to share the good work that people are doing, which would enlighten people. So like when we did our own symposium here about two months ago, I had seated only 10 projects two years ago. There were 300 people lining up to tell me what they’re doing in AI. Right? I don’t know who these people are. I never gave them any money. But it was amazing for these people to come and talk about what they’re doing with AI. The humanists talked about AI, the artists talked about AI, the AI people talked about AI. And they all had a different perspective and a different way of using it. It was so rich and diverse that people went back inspired. Yeah. So that’s my job as a coach of the steering committee, is to create a framework which allows us to think about how we as an institution are gonna do, uh, work, be working with AI, and for every individual to see themselves in the framework and customize what they’re doing so that they achieve what they want to achieve.

Taiyo: I think that’s such a beautiful way to think about what- Maybe in this context, coordination means for an, uh, for a system as complicated and as diverse with so many different points of views and perspectives like the UC. You know, we’ve talked, uh, in the past about how, well, sometimes the, the rollout of AI in higher education is, is sometimes framed, is characterized as the Wild West, right? And it’s kind of, um, suggests a kind of anarchic nature, uh- Right ... a little bit rules free or something like that. But I think what having a kind of policy that’s kind of liberal, uh, allowing people to experiment and, uh, try many- Yep ... different things, succeed and fail, right? I think it really does lead to, with good communication of course- I think it really does lead to good outcomes for an institution.

Chancellor Khosla: So that’s one of the things we are doing on campus. I tell people I want us to be the skunkworks for AI.

Sarah: Yes.

Chancellor Khosla: And... You know what skunkworks are, right?

Sarah: Yes. Yeah. I use that phrase all the time.

Chancellor Khosla: Okay. I said, “Look, I want rapid experimentation, rapid failure, and rapid, uh, movement to the next experiment,” right?

Sarah: Mm-hmm.

Chancellor Khosla: In academia, we are not used to that because God forbid if we say it’s rapid, then we confuse that with being thoughtless.

Sarah: Right.

Chancellor Khosla: We don’t assume or believe that something can be thoughtful and fast simultaneously.

Taiyo: Yeah.

Chancellor Khosla: For us, deliberative process, a slow process, is the only thoughtful process, and I think that’s a fallacy. We need to get to a point where we have to understand fast and slow is contextual. I mean, if I’m making a policy change, fast might be six months and slow might be five years. You see what I’m saying?

CHAPTER 5 [22:08-24:39]

Sarah: Absolutely. So speaking of something we need to move fast on, California has millions of workers who are gonna need to reskill because of AI, right? So most of them probably don’t want a bachelor’s degree and never will.

Chancellor Khosla: Yeah.

Sarah: So presumably many of them are looking to upskill right now, and so I’m wondering if AI-supported education could let places like UC San Diego serve more people who aren’t traditional degree seekers.

Chancellor Khosla: Right. I think that’s one of the things you might see as a change that might happen to higher education because… In higher education, we took pride in the fact that, uh, not everybody, but institutions like mine, that we are an R1 institution. We only deal with primarily with residential students. We don’t deal that much with commuter students. We deal with high-end research. We don’t give out certificates, we give out degrees. You see what I’m saying? So I think I can imagine, and remember, that’s why the extension was created in the UC. The extension was to take what we are doing and extend it to the communities around us. I think we need to rethink that because every institution might have to do, uh, would be a multi-scale institution, would have a four-year degree, a two-year degree, a one-year degree, a two-month certificate. And in some sense we do. It’s the business schools that do that right now in executive training, but most other departments don’t do it. But I think we might have to start doing that because first of all, it’s getting too expensive to spend four years, even when you need like six months worth of in knowledge to do your job. So we have to rethink how we are gonna be, how should I say, delivering education at different scale, different timescales.

Sarah: And I guess the hard part is figuring out what goes into those shorter programs in the first place. With AI changing so quickly, we don’t necessarily know what a faculty member or an engineer is going to need two years from now. So how much can the university prescribe what that upskilling should look like, and how much does it have to start with what people themselves are trying to do?

Chancellor Khosla: Yeah, but the problem is we, you know, being educators, we think we know all the answers. So we basically, uh, force our answers and our thinking on everybody else. We need to give more- How should I say? We need to attribute more intelligence to the student and the other people who are using our services. Yeah. They know what the hell they want. We should be listening to them a little bit more than we listen to ourselves.

Taiyo: Love that. Yes.

CHAPTER 6 [24:40-27:46]

Taiyo: Yeah. Speaking of the sort of student experience, I know that you’re a big believer in residential learning, right?

Chancellor Khosla: Yeah.

Taiyo: And that you’ve invested more in residential life than almost any other university leader in the country. You’ve also described AI tutoring as kind of being faster, cheaper, and more individually effective potentially. Is there a balance that you see as ideal between the sort of AI supported individual work- which I guess could technically happen anywhere, right? Partially obviating the need for, uh, residential infrastructure and the kind of in-person living and learning communities that you’re creating at UCSD.

Chancellor Khosla: So the in-person living and learning, at that age, what’s happening to you? You’re growing physically. Physically, you’re still not completely grown. That’s why, uh, the football team, uh, the NFL, doesn’t pick an 18-year-old. They pick, like, people at age 22 and 23. You see what I’m saying?

Taiyo: Mm-hmm. Quite right.

Chancellor Khosla: You are not completely emotionally mature at that age. You’re growing. You don’t know how to socially interact. I mean, you might have been clumsy with meeting, uh, the opposite sex or finding friends, and then age 23, suddenly you’re a better person at meeting other people. You see what I’m saying? Mm-hmm. Uh, you’re growing psychologically. I can just go down the list. There’s multiple growth phases you’re going through, different aspects of growth, and I think by being together, it helps those growth mechanisms. It’s not just knowledge. It’s not just learning knowledge. It’s also understanding how to deal with human beings. Most of you, like even as an academic, if you look at what it takes to be successful at a job, it is more what we call the soft skills than the skills you might have learned, because after three years they all become useless anyway.

Sarah: [laughter]

Chancellor Khosla: Especially in the professional areas like engineering, right? Mm-hmm. I mean, the field is changing. Like, uh, what I did for my PhD is right now an undergrad- homework project. So literally that’s how much the field has changed.

Taiyo: Yeah. That’s hilarious. Yeah.

Chancellor Khosla: You, you see what I’m saying? Yeah. Yeah. What was not known then, and I helped make it known through my research, now is not only known, it’s in practice.

Taiyo: I mean, that’s amazing. Yeah. That’s amazing that the work that you did-

Chancellor Khosla: No, but that’s the nature of science, right? Yeah. Yeah. That’s the nature of- That’s… even the science we... Like physics, you know, what was unknown and a Nobel Prize winning idea, now you learn in your undergrad course in physics. For true. So that’s just the nature of science. It’s also the nature of most fields, right? So the point is, what looked impossible then is the normal practice today.

Sarah: Taiyo and I talk all the time about what skills we put on that list. Curiosity is one that we put in there always at the top and we kind of wonder, can it be taught or is it innate? I think it’s a little bit of both.

Chancellor Khosla: I, I think you’re right, it’s a little bit of both, but I think if you’re a lazy person you will not acquire it that easily because being curious takes some work.

Sarah: Right.

Taiyo: Yeah. You’re right.

Chancellor Khosla: You see what I’m saying?

Sarah: So true. At the very least you gotta get up and look at what the thing is - the, the noise you heard.

Chancellor Khosla: Exactly, right?

CHAPTER 7 [27:47-32:48]

Sarah: So I guess following up a bit from what Taiyo was saying on the, on res life, I know you’ve argued elsewhere that university should measure itself by how broadly it delivers excellence, not how many people it turns away, and I think you used the phrase elite but not elitist. Right. And so what does an elite but not elitist AI-supported education look like, and who is it for? Is it for a different audience than you currently serve?

Chancellor Khosla: No, it is for that audience and many more that we don’t serve, that we turn away because we are space limited. Right? And I think most people confuse or think that elite and eli- elitist are the same, but they are not the same. They are two very different words. They might have the same roots, but they are different words.

Taiyo: Yeah.

Sarah: That reminds me of something I think Arizona State University’s president said at the Learning Lab INSPIRE 2024 convening - that education shouldn’t be measured by exclusivity. It struck me as very similar to your idea of being “elite but not elitist.” But ASU seems to have pursued inclusivity by massively scaling up their online and remote course offerings, while UC San Diego has pursued it through an expansion of in-person residential life while still remaining very selective and very invested in residential education. Are those competing models for the future of the public higher system?

Chancellor Khosla: They have to coexist. So here’s the way I think about this. The way I think about this is if I am offering something for free or extremely cheap, uh, where like very inexpensive, then I can just be in the mode of delivering what I’m delivering without trying to understand is it being absorbed at the other end or not? Okay? If I am charging some reasonable price to deliver what I’m delivering, then I think I owe it to you, you the people I’m delivering it to that the outcomes are proportional to what you are expecting from, uh, spending your time and money. Does that make sense?

Sarah: Yeah.

Chancellor Khosla: So the point is that as an institution like ours, graduation rates is a way to figure out, uh, are we making progress in our goals? Are the students making progress? But if I was just giving courses for free, then I could be like Coursera - just put it on the web, and then I’m done. There’s a place for both of these and other modes of, uh, delivery to coexist. We cannot say one is better than the other because the individual who needs it gets to choose which mode they want.

Sarah: Mm-hmm. Yeah. I’m thinking a lot about there’s, recently in the news, there’s been a lot of stories of people, um, it’s kind of the academic equivalent of token maxing, like they’re just taking, they’re enrolled in 20 asynchronous online courses and getting their degrees in months.

Chancellor Khosla: Right.

Sarah: You know, and this is like leading to this big crisis of, well, what does a degree even mean if you can accrue them that quickly? And what did you actually learn? I want us to think a bit about the future. Like what other changes do you expect to see in higher ed and mixed modalities? I don’t ever see, I think, in-person education going away.

Chancellor Khosla: No, it’s not gonna go away ...

Sarah: even though I see it being kind of culturally devalued in a lot of venues.

Chancellor Khosla: It’s, uh, yeah, I don’t think it’s gonna go away, but I think we will accept other modalities more easily than we have in the past. So I don’t know if you rem- you’re like probably too young to remember. There used to be something called correspondence courses. I don’t... Do you know what I’m talking about?

Sarah: I’ve heard of them, but I don’t know what they are.

Chancellor Khosla: Okay. There used to be these courses long before TV modality became rampant, where they would send you the course material by mail.

Sarah: Oh, yeah, and you mail back your stuff. Yeah.

Chancellor Khosla: And then you would do the reading, do the test, and mail it back. So that was a correspondence. You know, that was like using snail mail as a way of communication. Right? Uh, then there were courses on TV, then there were courses on the web. Yeah. So these modalities have existed from, uh, day one of, uh, education. It’s just that, uh, the speed of delivery changes and our ability, and we, whether we accept them or not. So in those days, if you got a BA using a correspondence course, you were inferior. [laughs]

Sarah: Hmm. Right.

Chancellor Khosla: You would not say that today. If you get a BA using online education, which is the same as a correspondence course, by the way, very similar you would not think that that is any, any inferior, right? Yeah. Now you would make a difference be- differentiate between for-profit and non-profit colleges, where for-profit would have lower credibility in your eyes than for, than non-profit.

Sarah: Right. Right. Right. Do you think that’s because the quality, the, like, were the correspondence courses high quality and it was just a cultural stigma against them?

Chancellor Khosla: It’s the same course that you’re reading in your textbook except it’s coming by mail!

Sarah: Got it: the mode of content delivery is not always a good proxy for academic quality.

CHAPTER 8 [32:49-41:31]

Taiyo: I mean, I think one thing that I’m observing as a faculty member, and I’m seeing students who are coming in with the credits from some online programs. As an instructor, I’m noticing that not in all cases, but there are significant cases where students are coming in, and they do not have what their transcript should suggest they have in terms of like the intellectual foundations and that sort of thing. And one suspects highly that there is some kind of credentialing issue here, that what the transcript is suggesting is just not what is actually inside of my student’s head in that moment.

Chancellor Khosla: So, yeah. Okay. So yes and no, right? So again, it’s a nuanced thing. So for example, is there grade inflation today? Yes. If you get A in arithmetic, I don’t... what does that “A” mean? I don’t know. It means different, different depending on where you come from. Now having said that, let’s just pick, uh, what am I gonna pick? I’m gonna pick some reasonable institution that’s ranked 50th in the US News and World Report. Now, if you take, let’s say a calculus course from there, and then you come to me, the likelihood that you remember everything in that course is close to zero. So now if I start testing you under the assumption that you remember everything, then we got a problem because I’m gonna find holes in your knowledge. But if I test you on the very basics, on the three things that you need to know, and if you don’t know those, then we have a different problem, right? And I think what you’re saying is it’s those three basic things that everybody should know if you’ve taken a course in algebra, that people don’t know. And I think that is partly true because of the way it’s taught, partly true because of the way people are tested, so they’re not tested exhaustively and people just give grades. In high schools, there is no real, how should I say, consequence of pushing a student into the next grade. In fact, there’s a consequence of not pushing the student to the next grade, which is not true in a college. If I start graduating kids who don’t know anything, none of my kids, none of my students will get any jobs. It will, uh, impact us completely. In high school, there is no consequence that I can see.

Sarah: So you got me thinking about the news I’ve been hearing coming out of UCSD about addressing gaps in math readiness, or quantitative reasoning more broadly.

Chancellor Khosla: I knew you were gonna bring that up.

Sarah Hey, you brought up, I was like, I gotta go here!

Chancellor Khosla: [laughter]

Sarah: But Learning Lab funded the expansion of an AI tutoring platform that folks at UCSD had developed with the math department - I think starting in early 2024. And I remember back in February, Taiyo basically came running up to me, all wide-eyed and excited, after he heard the preliminary results the team had presented.

Taiyo: Yeah, they reduced the failure rate for these gateway math classes by two-thirds.

Chancellor Khosla: Yeah. The, that’s the, that’s the one of the projects I started. Right. Exactly.

Sarah: Well, what struck me about that is that all the college math educators I know have been talking for years about students arriving without enough mastery of pre-algebra to succeed in college algebra, let alone calculus.

Taiyo: Yup!

Sarah: So it seems historically, addressing those gaps at the individual level takes an enormous amount of instructional time and support. Does something like an AI tutor change what’s possible there? Can we suddenly provide that kind of individualized support at a scale we couldn’t before?

Chancellor Khosla: I don’t even know if we have aggressively touted the virtues of that course, but I think that’s what’s gonna happen. Look, every generation is different than the previous generation, so we need to, as an educator, my view is there’s a class of people who made a decision on admitting these students. Once they’re admitted and we find out that they have holes in their education, in their background, which was not obvious during the admissions process, it becomes our responsibility to fill those holes and get them through a normal four-year education. Uh, I think we, if we wanna solve a problem, we have to deal with the high schools directly, but we cannot be blaming the students for not knowing what they were certified as knowing.

Taiyo: Yes. Very true.

Sarah: Of course.

Chancellor Khosla: And I also find that as faculty members, the more the age gap between a faculty member and the student, the more likelihood the faculty member has a - how should I say - a more romanticized view of how much they knew at that age.

Taiyo: Oh, that’s hilarious. Yeah.

Sarah: Taiyo and I talk about “golden age” logic all the time and especially the tendency to remember college before AI as though the whole thing was already calibrated to produce learning. But if you think honestly at our own experience, how often were we actually getting out of an assignment the assignment was supposedly designed to teach us? I mean, I probably did less than half the assigned reading in some of my college classes.

Chancellor Khosla: I didn’t even attend half the classes, you know.

Sarah: We were not angels ...

Chancellor Khosla: if you test me, if you test me on some of these courses, I barely know, like, 2% of what was taught, but I still passed- Totally ... the test with an A because I knew what to read.

Sarah: Exactly. Exactly. Taiyo loves this example that I placed out of college math, uh, on the AP exam for calculus, and I asked Taiyo, I’m like, “I don’t know what a derivative is. I just know you take the number here, you multiply it by this number here, you do this thing with it, and then you, you get it.” And he’s like, “Do you know what you were doing?” And I’m like, “Nope, just moving numbers around like a, like a brainless bot.”

Chancellor Khosla: Yeah, some of them, some of it can be mechanical. I mean, mechanically you can do things without understanding. And, you know, that’s a good example of, uh, both an integral and derivative. If you don’t understand conceptually what it is, you can still do it mechanically without knowing what the hell it means.

Sarah: Exactly. And our human assessment tests were not good at actually telling the difference.

Chancellor Khosla: Right. Exactly.

There was no written question saying, “What exactly are you doing here?” So I think, yeah, that part is so crucial.

Chancellor Khosla: And this is where I think AI is gonna make a difference. I mean, you as a human being, if you’re persistent and resilient, you would go to the AI and learn.

Sarah: Well, this is why I think Triton is so interesting, because it’s been calibrated so that it doesn’t just give the student an answer, right? Like, with our institutional accounts, I usually have to teach my students how to prompt it to make sure that it’s gonna generate some productive friction and not just give the answer, but that seems to be something that we can work out, or we can teach people how to use it in a way that is not gonna just let them offload their thinking. But then I think there’s this bigger question of uneven adoption, right? Who learns how to use them in ways that actually extend their own, you know, the person’s capabilities?

Chancellor Khosla: Right. AI is gonna be the greatest equalizing force, and it’s also gonna be the greatest force that’s gonna create inequities if we don’t deal with it properly. It’s gonna create a greater amount of inequity because it takes money to build agents. And the rich will keep on making more agents and getting richer, and the poor will have no money to do anything and keep on getting poorer. Because now you, you see what I’m saying? You will have inability to deliver the same work that you are able to deliver. I would have a problem, right? Mm-hmm. So I think we need to think about this, and this is one of the issues I wanna make sure that the steering committee deals with. What policies do we need to put in place? How do we make sure that the equalization happens better and inequities don’t happen at all? Is this making sense?

Sarah: Oh, it is. I, it just, I have to wrack my brain on what, what policy, at least under the conditions we’re in right now, I have a hard time imagining what would go into such a policy.

Chancellor Khosla: I don’t either. That’s why I’m putting it out there to you and your readers and your viewers, and if anybody has thoughts, just drop me a line.

Sarah: I’m gonna ask Claude after this. [laughter]

Chancellor Khosla: No, trust me, I’ve done that already. [laughter]

Sarah: Any insights?

Chancellor Khosla: But remember, Claude is trained on the same knowledge that, uh - everything deficient in a human being is in Claude. Mm-hmm. So to expect Claude to be a super god is a fallacy, and that’s the other thing I wanna say: We give AI the power of God, as in, like, omnipotent and I think it’s not the right way to look at AI. It’s got the same deficiencies that you and I do, except slightly less so because it’s trained over thousand versions of you and me, millions version. You see what I’m saying?

Sarah: Mm-hmm. We also like to say you can de-bias an AI faster than you can de-bias a human.

Chancellor Khosla: Exactly.

CHAPTER 9 [41:32-45:33]

Sarah: So we a few “rapid fire questions” we prepared for you. Let’s start with” If you were building a university for the first time in 2026, what would you reproduce and what would you not reproduce?

Chancellor Khosla: I would give every individual student coming in the ability to define what they perceive to be their general education. With guidance from some faculty, I think I would give people the ability to customize everything that they have, right? So that we don’t have hardly in any place right now.

Sarah: Mm-hmm. Be still my heart. I was a GE director and a GE chair for years, and that is my dream for gen ed.

Chancellor Khosla: Is what? To make it uniform?

Sarah: No, no, no, no, not uniform, but exactly as you said, you know, to, to make it so that every student gets a kind of their own agency in deciding what constitutes the general study that should best supplement their major study.

Chancellor Khosla: And now if you think about AI, we cannot do it because we don’t have enough money to hire all the faculty and give two-person classes. But now if you think about all of these faculty members and, uh, you know, uh, these capabilities in AI agents, so now I would like want to teach students how to use an AI to get the part of their education.

Taiyo: Yeah, even if you did have the money, an implausible percentage of the population would have to become professors to maintain a 2:1 student teacher ratio.

Sarah: Oh man, what a dream… okay next one:

Taiyo: You did your PhD in robotics in the ‘80s, so you lived through the AI winter, as they call it, where the whole field was kind of written off. So what does surviving an AI winter teach a chancellor about navigating the current AI mania that we’re living through?

Chancellor Khosla: I should have survived. I should have hung in there longer because then I could make money now. I could be working for Anthropic or OpenAI.

Taiyo: You could, yeah.

Chancellor Khosla: No, I mean, look, I, I had a great career. Even in the winter, uh, during the winter times, I had access to great resources. I had a great career. I was privileged, to be honest with you, coming from Carnegie Mellon, the hotbed of that area. Mm-hmm. So I was never short on resources, right? But I do think that it was a tough decade or two for several people, and they gave up, but I’m glad that it is back in business. You know, I never thought robotics would be a trillion-dollar industry but that’s what we’re talking about today. We barely had a million-dollar industry 30 years ago.

Sarah: Wow, that is wild. Okay, so you’ve drawn a distinction, I think, in some of the work I’ve read about you, between using AI for research and having AI do the research, so-

Chancellor Khosla: Right ...

Sarah: suppose an AI, like, can propose the hypothesis and design the experiment and control the robotic lab and analyze the data and draft the paper, what’s left for the scientist in that case?

Chancellor Khosla: So first of all, science is really, really, really broad. So for example, AI can do that in the context of chemistry much more easily. And there is a company, by the way, that is doing exactly that, that is using AI to exponentially increase the rate of discovery. So when I’m running, uh, you know, like in drug discovery, trying to figure out which drug will, uh, react with, uh, will, will be helpful for which, uh, disease, running, uh, essays, that can happen robotically very fast. But trying to figure out what’s the molecule, what’s the structure you want the molecule to... Like you s- So that can also be helped by AI, but I think there’s a human intu- there might be human intuition needed. Hmm. I can imagine in 10 years, AI will basically be like you and I, but I don’t think it’ll ever replace you and I. It would always be an amplifier of our abilities. Uh, the combination of a human and AI would be much, much more powerful than just an AI.

Sarah: Right. And then the question becomes how we build institutions that actually support that combination.

Taiyo: That’s a great place to end.

CHAPTER 10 [45:33-55:52]

Sarah: So, Taiyo, I was thinking when I was reviewing our interview footage about the overlap between Chancellor Khosla’s attitude about institutional innovations and your whole let a thousand flowers bloom approach of, like, giving lots of individual educa- you know, educators in their different fields this chance to experiment and to use their best judgment that they have as domain knowledge experts in their fields and seeing what works.

Taiyo: Right.

Sarah: Even though they’re all, or many of them are new to AI. So that makes sense to me, you know, partly because I think you and I both thrive when we have very little oversight. Like, we clearly share this kind of personality disorder of preferring very- ... very few constraints. Or at least, you know, the freedom to explore lots of options for ourselves without intervention.

Taiyo: Wait, wait. No. No, no, no, no. That’s, that’s not a disorder.

Sarah: Tayo, what happens if somebody tells you you can’t do something? Like, does it make you want to do it more or less?

Taiyo: I mean, okay. I mean, maybe marginally more. Mm-hmm. But again, not a disorder.

Sarah: Okay, wait, wait, let me rephrase. What... If somebody tells you that something can’t be done, are you gonna work around the clock to find a way to do it just to prove them wrong?

Taiyo: I mean, I don’t know, Sarah. I mean, I like to think that my executive functioning’s ability to delegate a- agency toward a goal is proportional to how much I actually want to achieve that goal regardless of external commentary, right?

Sarah: Yeah.

Taiyo: But, you know, as soon as I say that out loud, I have to be honest, it kind of strikes me as delusional.

Sarah: Oh, good. Know thyself.

Taiyo: I mean, so, like- you know, even though there is something fun about being right- ... it’s even more fun to prove somebody else wrong. Can we agree with that?

Sarah: [laughter]

Taiyo: And you know what? When I say that out loud, you know what? Maybe that is a disorder.

Sarah: Thank you.

Taiyo: But, but honestly, like, the way you’re describing it, it just makes us sound purely spite driven.

Sarah: Oh, no. No, no.

Taiyo: And, and yeah, and I don’t think that’s it. Like- Right... for instance, you know, Sarah, I love you. You’re my friend. No spite at all.

Sarah: Yeah.

Taiyo: But also, I love proving you wrong.

Sarah: Same! I believe you. We, we work well in teams. You know, we work- Right... with a lot of people, and they seem to wanna keep working with us, so I don’t think we’re spite driven. But, and I think we’re gracious losers when we are proven wrong sometimes.

Taiyo: Um... okay. Maybe.

Sarah: I mean, I guess the point I wanna make is that maybe we’re biased, right? Like, we, first of all, have a very high tolerance for existential dread. Hmm. And then... we have a very strong preference for a lot of autonomy, and so I think I’m just trying to be self-aware about why I am so compelled by this vision.

Taiyo: Yeah, it’s true. And you know what? I, I won’t speculate on the reasons we’re built like this. But I totally agree that we seem to be, shall we say, out of distribution in our appetite for autonomy.

Sarah: Anyway, I think that’s one reason I suspect that we’re both drawn to Chancellor Khosla’s point that mandated coordination is not a good approach. I think his phrase was to, it’s not a good approach to change in complex systems. Right. But you were both, I think, against chaos, right? We’re not like, “Yeah, everyone should carelessly try something, like yolo, throw caution to the wind,” impulsive about stuff, right? Yeah. But I think that we realized, you know, as soon as we started learning about the capabilities of AI, we were like, immediate action is better than slow de- deliberation in this case. And that’s because nobody knows what works pedagogically with AI yet, and the space is very large, and I think a decentralized search under genuine uncertainty beats a top-down mandate central planner who has, like, very little informational advantage.

Taiyo: Yeah. I think that’s true in a situation where, yeah, you have a high level of uncertainty, the task is about discovery.

Sarah: Right.

Taiyo: And you can keep the good stuff and sweep away the bad stuff, right?

Sarah: Yeah, exactly. So we’re still in the discovery phase. Results of studies are starting to come. We’ll see. I’ve been thinking about his other comment about skunkworks and how I love a skunkworks.

Taiyo: You love a good skunkworks.

Sarah: Yeah, I love a good skunkworks, right? But skunkworks need a lack of constraints, right? And I wonder if the faculty who experiment are disproportionately the ones who have fewer constraints. Oh. I mean, I’ve been reading a lot of ... There’s a lot of stuff in my news feed lately about, like, it’s back to school season, so here come, like, the doomer articles about how all these professors are retiring rather than adapt to AI. But I actually ... I mean, I’ve seen a lot of in- like, innovative work happening with senior colleagues who, right, like us, fully tenured and promoted, have the privilege to take a risk on, like, jumping on something unknown and devoting time to learning about something that isn’t ... that is new. I think that’s not distributed evenly, right? And especially during a formative period, where there’s all kinds of stressors and things. So I love a decentralized regime. But I also think about, you know, what’s left out of that.

Taiyo: Right. Right, right. So it’s, it’s kind of like when you have this decentralized regime, it will distribute or tend to distribute advantage along, along resource lines that already exist.

Sarah: Yeah. Possibly. Possibly. Tentative hypothesis. I wanna learn more.

Taiyo: Interesting.

Sarah: The other thing that really stuck with me about this interview too was the point about expanded access and his points about continuing education essentially, right?

Taiyo: Mm-hmm.

Sarah: Universities are not historically very good at coming up with a just in time curriculum, right, for massive disruption.

Taiyo: Yes, I totally agree. And, the disruption that AI represents on our economy presents a tremendous opportunity for institutions of higher education to serve a role in re-skilling and upskilling the labor force for a new economy.

Sarah: Right. Exactly. And I think, you know, there’s very good reason to say we want to make sure that the curriculum we’re building is the best curriculum and evidence-based curriculum, but we are in this very strange interregnum, I’ve called it before, where the need is there before we have the time to make really, really well-informed decisions.

Taiyo: Yeah. I think it’s interesting that these ideas came up in a conversation with the chancellor of a UC, right? Because we know that we often talk about these types of things at the CSU level and we know of course, that the community colleges are thinking about this as well. This feels like an enormous opportunity for all of these segments of California public higher education to come together and really think carefully and critically about how we want to execute on this potential for upskilling and reskilling, um, California workers.

Sarah: Right. We mentioned at the beginning of this episode INSPIRE 2024, and I think one of the most impactful conversations I had over lunch was with a community college colleague who I, I hadn’t met before, and they taught writing as well, and we were talking about all the types of things, you know, we were thinking about experimenting with - this is in fall 2024, and one of the things they said was that they felt really constrained as a community college professor by the need to articulate courses. So for anyone outside of academia listening, basically the need for a course to transfer in and count for a course that they’re getting in the CSU. And what they were worried about is if they really wanted to put in a huge AI literacy component in their intro composition class, but if that’s not something, if then, if that means then downgrading the amount of time spent on, like, grammar and mechanics, there’s a risk that some of the other courses in the rest of the system would not accept that, and then you’re adding added expense and stress to a, a student’s life, right?

Taiyo: Mm-hmm.

Sarah: I guess the, my point is that, you know, a failure to push people to innovate back in 2023, like a f- like, the failure to do it what UCSD did and say, “This is a problem, and we are not sticking our heads in the sand. We are on this, and we’re not gonna be on it by making, like, a ban on AI or something that was clearly gonna be completely useless.” Right? But they got on it in a way of let’s actually bring members of the university, including the, I believe they were students or alums, who built this thing. So anyway, I think that’s important that a big R1 like UCSD is setting that tone, right? When some- when a university that prestigious does it, it sets the expectation, or I think it normalizes it, right? It’s, like, culturally weighty.

Taiyo: Oh, yeah. Massively so. Yes, absolutely. They are leading by example, let’s say.

Sarah: Right. Thanks for listening. This has been my Robot Teacher. I’m Sarah Senk.

Taiyo: And I’m Taiyo Inoue.

Sarah: You gotta say your usual shit.

Taiyo: Oh, uh, it, if, if you enjoyed this conversation, please share it with a friend or with a colleague.

Sarah: Every time you say this it gets weirder.

Taiyo: I don’t care. Share it with a family member.

Sarah: “Send us an eeeeeemail.” That’s my favorite.

Taiyo: We love when we get email from you all.

Sarah: We actually really do. We do. We got some really nice ones this week.

Taiyo: Oh, yeah.

Sarah: It made me very happy.

Taiyo: Yeah, exactly. And any feedback that you can provide, positive or negative, would be much appreciated. See you next time.

Guest:

• John DeNero: Faculty Director of Data Science Undergraduate Studies (DSUS), Giancarlo Teaching Fellow, UC Berkeley Department of Electrical Engineering and Computer Sciences

Also available on: Apple / Spotify

Chapter 1 [0:00-4:31]

John DeNero: My guess is that we’ll discover that, like, a college education is maybe even more valuable and more important than it was before.

Taiyo: Welcome to My Robot Teacher, a podcast about how AI is reshaping higher education, hosted by two professors inside the CSU, the nation’s largest four-year public university system. I’m Taiyo Inoue, Professor of mathematics at Cal Poly Maritime Academy.

Sarah: And I’m Sarah Senk, Professor of English at Cal Poly Maritime Academy, and that voice you just heard a moment ago was John DeNero, the Giancarlo Teaching Fellow in the UC Berkeley Department of Electrical Engineering and Computer Sciences. John’s an expert on natural language processing and computer science education, and he teaches and helped develop two of the largest courses on Berkeley’s campus: introductory data science and introductory computer science for majors. And these are lecture courses with literally thousands of students in them at a time. If he sounds a little tentative at the start of this interview, that’s because we basically ambushed him at the National Workshop for Data Science Education right after he brought down the house with a comment about how - even though he’s optimistic about the positive potential for personalized support for students - that right now, “the kids are NOT okay,” as he put it, because for the past 4 years or so, half of them have been just having ChatGPT do their homework for them. And he agreed to come into a back room with us, and huddle around a microphone for an hour talking informally about what it’s like to be a computer science professor in the age of AI.

Taiyo: Yeah, I think this was a really wonderful interview, and it’s easy to see why John DeNero’s an award winning Berekely Educator. And just to note: this conference - it’s very relevant to my work in data science education, but Sarah was crashing the conference. I mean I think this is a tradition now at this point because I think this is the third year in a row that you’ve sort of shown up.

Sarah: The very first time you texted me, I think, saying, hey, what are you doing tonight? Want to get a drink at Berkeley? Like, meet me here. And I show up and there’s like 100 people with badges. And I’m like, Taiyo, is this a conference?

Taiyo: I kind of ambushed you, I guess. [laughs] But yeah, I think this interview is wonderful because we discuss some really important and pressing questions for higher education. And in particular, we get to talk about it through the lens of computer science, right? So in particular, we can think about how nowadays AI we know can write code and it can write code really well and perform all sorts of different technical tasks. So the question now becomes what still remains worth teaching? Is a computational way of thinking still worth learning in 2026? And then, we can Zoom out then a little bit, and ask what kinds of - or what elements of - AI education should every student in the university receive? And if AI can do students homework, what should homework BECOME in the age of AI?

Sarah: Yeah, for sure. Right. And, you know, I think as we’re thinking about how to scale up access to support and support that is accessible outside of the structured time that you, the expert instructor, have with students. For me, the question is how you instill in students a habit of the mind where they don’t reach for help of any kind too quickly. And they’re not bypassing the productive struggle that is sometimes uncomfortable, but I think is a really core part of learning and retaining something long term. And so if an AI tutor maybe helps them get to the answer a little too quickly without them thinking about it, or if they’re not training themselves to resist the urge to be like, help me right away before I’ve actually struggled through it myself, I think that’s a big problem. And I’m not even thinking about it in terms of academic honesty here; I’m thinking about it in terms of their own cognitive development. And I think that is the core question that John helps us think through: How do you preserve those elements of cognitive development when right now, at least, LLMs are not yet perfectly optimized to hit just the right amount of productive friction to help a student learn, but not wait so long that they think “this isn’t for me, this is too hard.”

Taiyo: So thanks again to John DeNero for doing this interview with us. And thanks to the organizers of the conference for allowing Sarah to crash the conference for the third year in a row.

Sarah: Much appreciated. See you all next year. I’ll actually register, I promise.

Taiyo: And as always, thank you to the California Education Learning Lab for sponsoring this podcast.

Chapter 2 [4:32-8:45]

Taiyo: What would you like the world to know about AI, John?

John: Well, yeah, since I mostly work on AI and education, I think, um, it’s worth knowing that the immediate effects of AI in education probably won’t be the long-term effects on AI education. AI was unleashed on the world in 2023. It immediately changed education. Like, every student started trying it out and figuring out how it fit into their life. Courses and universities and instructors tend to adapt a little bit more slowly than students.

Taiyo: You don’t say!

John: So we’re still figuring that out, yeah. So, you know, that’s just my perspective. It’s like, it’s, it’s now not something we can turn off. The world has changed, so we can’t, like, somehow go back to pre- uh, ChatGPT days, and it’s important to study the impact of what, where we are now, but I also just don’t think it’s very stable. Like, I think, uh, a lot of instructors, myself included, are changing the way we teach all the time in response. It’s, like, very dynamic, and we’re gonna learn what happens in the long run, and I guess I’m, I’m hopeful that in the long run, AI will actually help education. I just don’t think we’ve gotten there yet.

Sarah: So I’m curious, what makes you hopeful? What kind of capabilities or, um, context are you thinking about that’s motivating that hope?

John: Well, I think the capabilities we see out of AI are pretty impressive. Like, they, they can do a lot. They can provide the kind of, like, feedback to students that we couldn’t achieve for a long time. You know, we tried before this technology really matured to automatically give students guidance or feedback, but it just wasn’t, it wasn’t there. Now the capability is there to, like, say useful stuff to students. So that’s great. The resources required in order to provide this kind of input are a lot. Like, these things use a lot of energy. They cost some dollars, but they’re not nearly as much as, like, providing a personal tutor or you know, preceptor or something like that for every single student. Like, this was something that you can only do for, you know, princes of the kingdom or whatever, and like, yeah, you know, we’ve always been playing these games where we try to, we try to help a lot of students with one teacher, and, and it’s always been difficult. So, you know, there’s, there’s a chance that we’ll address that issue.

Sarah: I was gonna ask as a follow-up, when you say, like, we, we have ... we didn’t always do that, you know, you teach classes at Berkeley that have literally over 1,000 people in them, right? Is my understanding. Yeah?

John: Yeah, yeah.

Sarah: So that’s one aspect of it, but I think this tension between the ... or the ... When people talk about how resource hungry AI is, I don’t often hear that compared to the resource-intensivity of super personalized education for everybody.

John: Yeah, I mean, I think we only have so many adults around, and, like, we’d have to use a lot of them to teach children if we wanted to have, like, an adult mentor for every child all the time. Like, it’s, it’s hard to achieve. So that’s always been the kind of resource constraint we’ve dealt with, and it doesn’t look so bad. You know, we have AI tutor prototypes running in our courses, and they cost, you know, $10 or $20 a semester or something like that in terms of resources. Like, that’s very different than giving a personalized tutor to somebody for a whole semester. So like, there’s - you know, you can’t measure everything in dollars, but it sometimes helps to gauge the order of magnitude of what we’re looking at. So at the same time, AI doesn’t provide what a human teacher provides…. yet. But, you know, my background’s in this technology. I’ve been working on it for a long time. I was very surprised by how well it became and how quickly it became… But I also don’t think that there are, like, clear limitations to where we are in the technology that would prevent it from being even more useful than it is. I think we’re gonna continue to see some progress, and that progress seems like it could, could matter for education. Like, uh, these things will do better at providing kind of useful feedback or input to students.

Chapter 3 [8:45-12:56]

Taiyo: I think there is a kind of fear that if we let the, quote unquote “machines” take over on the sort of feedback portion, the way that we interact with our students from traditionally what the instructor has had to do in, in talking and, and, and getting to know their students. There’s a fear that if that very human relationship gets replaced by a machine, that there’s something about that learning in- that the, the teaching and learning that is diminished when you remove the kind of care, for example, that presumably many instructors have for their students - the premise being that machines are incapable of a similar type of care.

John: Yeah, I mean, I totally agree. Okay. I think that machines are kind of a supplement to the human teachers. I don’t want human teachers to go away. I think we should probably have more of them. But the thing I’ve learned in my career is that the fact that I have a very large class doesn’t seem to be a huge problem. Like, there’s one of me and 1,000 students. How can this possibly work? Well, I’m not alone. Like, I have a bunch of TAs and tutors that help me teach.

Taiyo: Yeah.

John: And they have great human relationships with the students. And we’ve been trying to shrink the size, the, like, student-teacher ratios there consistently, and it seems to have always mattered, like going from 35-person sections to adding, you know, like five or six on one, uh- Hmm ... group tutoring and, you know, drop-in one-on-one help. Like, you want that interaction to be small. But the other thing is that going through the pandemic and then returning to campus, I came to believe that having a campus and having students, like, together, in a physical environment was really helpful to their learning, to their development, to, you know, them being excited, like, being engaged, knowing what they were supposed to be doing. Like, that’s really valuable. So I, you know, I don’t think either of these should go away. Like, we should have people teaching other people. We should also have students together, kind of learning together, working together. Mostly, I think about what students learn when they’re on their own.

Taiyo: Yeah.

John: Like, there is a lot of learning that happens, when you go home and do your homework and do a project. Or maybe you’re with somebody else, but, like, really, at some point, you have to think about something by yourself for a while and, like, try to apply it or write a paper by yourself or whatever it is. And that, I think, was Not, never a terribly well-supported process- Right ... like in our education system. Mm-hmm. Like, and so, you know, some students thrived anyway. They, like, figured out how to get a lot of value out of their homework. Some, you know struggled through it and, and were frustrated and were miserable and weren’t learning as efficiently as they could. Like, that’s the opportunity that I think, for applying AI. So it really has nothing to do with what happens in classrooms. Like, in classrooms you should have a class. You should hopefully have in-person classes, and students should work together and discuss and all that good stuff. But when they go home and do their homework, it was never perfect. It’s way worse than it was before because of public AI tools that students use as a crutch to get through their homework, and therefore don’t go through the same amount of productive struggle that they used to. I mean, some students do, but on average, I think students are reaching for help too often, and the help they get isn’t really the most constructive. Like, it certainly gets them to the answer to their homework, but it doesn’t get them to, like, think about it and have that learning experience, and, like, explore wrong answers before they get to the right one. You know, all that good stuff.

Sarah: Right.

John: But it could be better than it ever was. Like, having some simulation of a great tutor kind of walking you through things when you get stuck, but also probing your understanding so that even after you finish a problem, it makes sure you really understood it. That could be better. Uh, so that’s what we need to build. We need to build it pretty fast because right now it doesn’t exist and, uh, yeah, and students are kind of doing their homework - not in the most productive way. So, yeah, that, that’s where I think the, like, positive hope comes from.

Chapter 4 [12:57-17:02]

Sarah: Well, so what, uh, like, what solutions do you imagine? And, is it a technological solution, a human solution, both?

John: Yeah. Well, um, I’m looking forward to running more experiments to know what really works. Um, I think we’ve already seen that if you provide a companion with your homework hat gives you some indication of what you’re doing wrong when you get a problem wrong then students will use it. They’ll use it- Often in lieu of public tools that would just tell them the answer. You can give them a system that doesn’t tell them the answer, and they still like it. Like, we have a system that tells you, like, which paragraph in the textbook you should read in order to figure out how to solve this problem. You know, in programming assignments it tells you what line of code is causing the problem or walks you through the process of where things went wrong. I, like, I think all this stuff, I would have no problem having one of my human tutors give that kind of feedback to a student, so why not have the AI do it? And the AI can do it, like, much more conveniently.

Sarah: I was thinking a lot about… so the comment you made yesterday, and we’re at the Berkeley Data Science, I’m gonna botch the name ‘cause I’m crashing this event.

Taiyo: The National Workshop for Data Science Education, I believe- is what it’s called.

Sarah: Yes. Thank you for letting me crash.

John: You’re most welcome.

Sarah: Um, but I was thinking about the social element. Like, in my own, thinking back, and I - you know, we always say this: “We became professors. We are not good models for students, like, in general, we shouldn’t be, like, using ourselves as examples” but I’m gonna do it anyway and say that my fond memories of doing homework were, like, struggling through something myself and then not understanding it, and then calling, ‘cause I’m that old, a friend, and being like, “Hey, I know you’re working on this too. How’d you do this one?” Or, “What, what...” Like, and then there’s, like, a camaraderie and them being like, “Oh my God, it’s so hard too. Like, here’s what I did, and I couldn’t get it either,” and then we talked together. And then there’s a kind of generative thing that happens from that human conversation. I wonder how much of that is central to, like, students’ kind of cognitive development in ways that we haven’t really studied or that we don’t sufficiently understand. Like, maybe not even the development of social skills, but actually cognitive skills that come from, um, talking with somebody who also has a lot of, like, let’s say other baggage that maybe a machine doesn’t have, you know?

John: Yeah, totally. I think that there used to be, you know, big groups in certain dorms that would just like get together when the homework was due, and they would all mostly do their own homework, but they would like talk to each other and get advice, which I thought was fine.

Sarah: Yeah.

Jon: We used to have, you know, uh, scores of people come into our drop-in help sessions near deadlines. And they would, you know, talk to each other because they were waiting to talk to the TA, and there was a queue and all this stuff. It disappeared.

Sarah: Yeah.

John: I, I, I’m not sure that it disappeared, but this is my sense is that like this introduction of this technology changed the culture among students- Right ... where it was kind of expected that you would, like, go somewhere to do your homework, and like, meet up with people to do it. And now over the last few years, that’s not so much a thing anymore. I’m very tempted to try to encourage it again, like run some kind of s- you know, study hall, where we don’t force people to go to study hall maybe, but we do, uh, make an option. Like, oh, here’s a, here’s the place where everyone’s gonna be working on this assignment. Um, or we used to call them homework parties, and we’re gonna, we got a new building coming. We’re gonna do some of that in there. Like I, I, you know, I, I think it’s good to encourage it, but it’s also just reality that like the culture’s changed, and we don’t really get to dictate how students behave as professors. Like, if they’re not gonna meet up with each other in the dorm to work on it, maybe that’s too bad. Maybe it’s worse for them in the long run, but it’s… it’s not something that we can necessarily force, and you know, that, that’s happened independently of anything that we did. You know? That’s just ‘cause the technology existed. You could talk to your computer now. You couldn’t talk to it before. So I think the best we can do is encourage it, plus make sure that when they’re talking to their computer, they’re at least learning as much as they can per hour.

Sarah: Yeah.

Chapter 5 [17:02-22:30]

Taiyo: Yeah. You mentioned productive struggle earlier, and as we know, technology can allow us to circumvent some of that struggle, right? It allows us to offload some, some things that we just don’t want to do, and if it’s sufficiently easy to use, that can be a huge boon to our time and our energy and all of those kinds of things. How do you draw that line between the productive struggle that you want your students to be partaking in versus... Like I, I would guess that you’re probably not going to require students to be able to multiply five eight-digit numbers by hand. That would be a ridiculous ask in this day and age, ‘cause of course we would always go to our computers or our calculators to do that kind of task.

John: Yeah, it’s been ridiculous for a little while.

Taiyo: A little while now, right? Like, obviously then there’s some kind of feedback loop between the technology that exists, but then also the sort of learning outcomes perhaps you wanna call it, or maybe another phrase that I know you’ve used in the past, but a computational way of thinking. How should we think about these things, particularly as AI just gets better and better and better and better at doing a lot of these kinds of tasks?

John: Yeah, I think that there will be things that we de-emphasize and things that we increase emphasis on. Even when it comes to calculators, for the most part you want students to have some idea, or any people, not just students, an- some idea of what is happening when you do anything.

Sarah: Right.

John: Um, so you know, we still teach students the multiplication and addition algorithms. We don’t do it with eight-digit numbers ‘cause it gets boring after a while-

Sarah: Sure does.

John: and, and humans aren’t that good at, uh, carrying the, the one. But, um, but it does more than just show them, like, here’s a way you could do it. It also, like, builds some intuitions about what the result should be like. Like, if you spend some time multiplying numbers on your own, then you get the sense that like, you know. three-digit number times a three-digit number is gonna be a five-digit number. It’s not gonna be a 12-digit number and it’s not gonna be a one-digit number. You know? Like, this, of course you could, like, prove this mathematically, but I think it also just, like, becomes, part of your intuition about what’s going on to have gone through those skills. And you know, I teach a lot of programming. I view it similarly that- In order to use these AI tools effectively, it helps a lot to be able to imagine what the output of your request will be. So you wanted to write some program. In order to describe the program that you want, you have to kind of think a little bit, like, “Oh, what am I asking for?” And in order to specify it well, you usually have to think about what the outcome is gonna look like. And that’s why people who are good programmers are getting much more value out of, like, AI programming than people who aren’t programmers. Like, uh, I, I know, I know people- ... of both varieties, and it seems to be that the people who could have programmed this on their own, just would’ve taken them a month are flying, and they’re getting so much done.

Sarah: Totally.

John: And they’re, you know, correcting the problems that show up, and they’re, they’re making progress. And the people who never learned to program would very much like an app, and they describe it, and some app comes out. But then eventually they hit some wall where they don’t understand what, what happened, and so they don’t even know what to ask for, or what to describe or what to look for. And, you know, will the AI system fix this for them? It’s possible. Like, just because like a great human engineer can work with somebody who doesn’t know how to program and eventually, like, build what they want, but they do it with a lot of back and forth and a lot of probing and it’s, you know, it’s kind of a clunky process. Whereas instead, if everybody understands what, what’s getting built, then it’s, like, much more streamlined. Yeah. So I think there’s a lot of value in understanding, like, how computers work, what programming languages are, why we write programs in these programming languages instead of writing them in English, how that allows computers to, like, run the same program over and over again, get different results based on different inputs. Like, this whole world of programming, like, if, if you don’t have the right mental model for it, I just don’t think that you’re gonna be able to... You get as much value out of the AI tools.

Taiyo: For sure.

Sarah: I mean, we’ve had a lot of conversations about this, about, like, computational thinking, mental models essentially. Uh, the, and whether, like how much drudgery you need in a particular discipline to have that kind of innate, or you used a phrase like, um, that int- intuition about the way to do something. When I first like, experimented with vibe coding, as somebody with no computational literacy, I was just like-

Taiyo: Oh, I know what you’re talking about

Sarah: …brain dump, here’s what I want. And I didn’t give a lot of thought to, like, how to structure the stuff. Mm-hmm. And then in a lot of experiments and, like, having Taiyo with his m- you know, intermediate knowledge of Python.

Taiyo: Yeah. Yeah. I think that’s fair to say.

Sarah: Adequate knowledge of Python say, “Well, that’s not gonna work because you’re not scaffolding the steps, and think about it like you would think about teaching a student. What do you have to have them do first before they can do this step next?” And then from there, the outputs just, like, radically changed.

John: Mm-hmm. Um, yeah.

Taiyo: I, I think the, I think the anecdote that I use, which I think is- Been proven to be apocryphal? I don’t know. Is this, is this, um, thing about John von Neumann- being or somebody was saying, like, “Computers will never be able to love,” and John von Neumann saying something like, “Describe to me this love in sufficient detail, and we can get a computer to, to do it.”

Sarah: [laughs]

Chapter 6 [22:31-26:09]

Taiyo: I mean, yeah. I mean, I’m, I’m wondering actually if... Okay, you know, you are a computer scientist, and we have a humanist here with not an intermediate level of Python, very little experience with computer programming. I wonder how, how you would describe computational thinking to somebody like Sarah.

John: Computational thinking is a kind of collection of problem-solving techniques that help you manage the complexity of software programs by kind of using the same ideas over and over again, kind of like building up an inventory of, of ideas that help you kind of reason about some piece of software, even though you can’t, like, read it from top to bottom and understand every detail that’s happening at the same time.

Sarah: Mm-hmm.

John: ‘Cause, you know, that’s what computers are for is keeping track of all the details. Like, you, we don’t have to do that, but we do have to be able to create the software or describe the software, and so we need to know, um, how to, like, manage all that complexity. And, you know, a lot of it is stuff that is not typical of how humans do things. Like, it’s often the case that computers will solve a problem by just trying all possibilities, and seeing which one works, like this massive guess and check.

Sarah: Yeah.

John: Whereas humans tend not to want to, like, guess and check their way through life. They, like, think about what might work and then try that first. So yeah, so part of computational thinking is just getting used to this idea that you could do something that, like, seems really dumb, but if you did it enough times, you would eventually find the answer.

Sarah: Huh. Yeah.

John: And your computers are fast, so, you know. Like, how do you find the square root of, you know, 121? Like, one way is to do some math. The other way is just to try squaring every number until you get, until you get the answer that you want, you know?. And, and so, like, you’re part of, uh, th- these, these things that are sort of computational approaches that are different from how humans solve problems-

Sarah: Right

John: I think you just have to learn. Because we’ve discovered them over the years. Different computer scientists and mathematicians have discovered this big inventory of different techniques that work well, and, uh, and now we all have to know them. Kind of common vocabulary for how to write programs. But then the other part is stuff that’s very human, but you wouldn’t necessarily think to apply it to computers. I talk a lot about abstraction.

Sarah: Yeah. Yeah, you mentioned this yesterday.

John: Abstraction’s this idea that when you think about something, you don’t need to think about exactly how it works, you just think about what it is and what it does. It’s, like, from one of my intro lectures, but you’re talking to me, you think, you know, John’s describing abstraction. You don’t have to worry about, like, how many oxygen atoms are in my body right now. That’s, like, not your problem because you’ve abstracted that away. This is something that humans do, is that they kinda give names to things, and they describe what they do without having to worry about all the details of how it’s done. And that is extremely helpful in managing the complexity of software programs because software is always built out of different parts. And those parts have extreme detail, but sometimes simple descriptions. And so as long as you can think about the different parts according to their simple descriptions, then you can figure out how to plug them together and build more interesting stuff. But exactly what that looks like in practice means that you have to actually look at some computer programs and, and work with it for a while. So, you know, that’s where you’re heading, is you’re just, like, using your own human ability to abstract away the details of things and learn how to do that when you’re talking about software instead of, like, uh, stuff in the world.

Sarah: Yeah.

Chapter 7 [26:10-34:32]

Taiyo: So a magic word here is compositionality.

Sarah: Mm-hmm.

Taiyo: And the reason why I mention that to you is ‘cause we’re gonna- we’re gonna interview in the future, or not so distant future I hope, um, a category theorist who’s gonna really, I think, drill into the idea of compositionality. I mean, that’s in a typically a mathematical context. But this idea of abstracting, of, of having, you know, these, these parts that are highly detailed but then can fit together and talk to one another, you know, that this is the, at the heart of compositionality and of computer programming, I think as well. Yeah.

Sarah: I mean, that seems to me also, like, a fundamental of human communication, that it’s based... That abstraction is sort of, like, the concept version that, you know, that maybe all three of us would share when you hear the word “software” But all three of our abstractions might be slightly different even though we’re using the same word. And a lot of, like, trust is built, I think, on recognizing that, oh, the, the way you’re thinking about that abstractly is also the way I’m thinking about it abstractly.

John: Mm-hmm.

Sarah: Often in class I’ll realize, like, misunderstandings when I say something and then I see something that a student’s written in a free write about what today’s class was about, and it’s totally different than what I intended. Like, how is it different with, from human communication and when you’re giving a computer instructions or you’re building a computer?

John: Yeah. No, I mean, you’re, you, you hit it, one of the core things that you have to learn in computer science, which is that programming languages are different from natural languages like English in that they were designed not to have ambiguity.

Sarah: Right.

John: They’re much less expressive. Right. Like in English, we can really talk about anything at any time at any level of detail. Programming languages are kind of much more restrictive in what they can describe, but there’s no ambiguity. They’re a constructed, artificial language, where when you write down the program and you run it on this computer or you run it on that computer, it’s gonna do exactly the same thing. We call it an interpreter to run the program but there’s no subjective interpretation. It’s always, like, an exact meaning corresponding with every sentence. Okay. Wouldn’t that be nice in English if there was an exact meaning corresponding to every sentence, but it’s not.

Sarah: That’s the dream!

John: Yeah. That’s not the case. That’s not how human languages work. It’s that there is this subjective element of interpretation, and that just doesn’t exist in programming languages. That’s what makes everything, like, run reliably.. Like, you know, the same app runs on everybody’s phone, so it’s good. But what’s bad about it is that all of a sudden you’re using this language that’s very different from English, where in order to express yourself, you’re constrained to use exactly the meanings that are built into the language. You don’t get to, like, add your subjective interpretation, and so you- Right, right ... have to kind of work your way around this constraint, and that’s part of programming.

Taiyo: That’s... Yeah. Wait, I... Can I ask you a question?

Sarah: Yeah, yeah.

Taiyo: Okay, I’m gonna make it a... Instead of a question, I’m gonna make an assertion and see what you think about it. I think you like the subjective element of natural language.

Sarah: Yeah!

Taiyo: I think that if, if nat- if natural language were - only had one interpretation, I think that you would find that to be an impoverished state of affairs.

Sarah: Oh, that’s really... That feels very right.

Taiyo: [laughs] Okay.

Taiyo: I had a feeling.

Sarah: Yeah. I do feel that.

Taiyo: I had a feeling.

Sarah: But I, but at, at the same time that I say I feel… So here’s why I say: that would feel to me like an impoverished state of affairs. It would, like... It feels like… if we were talking about this, you know, plenitude where words and meanings are, like, coinciding exactly, and not, like, mutually reproducing and cycling and context-dependent, I think there’s also a sense of - you know, I said that’s the dream, right? - Like, that’s where meaning would be kind of self-evident or would be... There would be, like, perfect, pure communication or something like that. That seems to me like a world that’s no fun.

Taiyo: [laughs]

John: Yeah, no, I’m with you!

Sarah: It’s a world with no drama. It’s a world with no, like, ability to exploit language’s ambiguity.

Taiyo: Yeah, yeah.

Sarah: For fun, not for evil! For like art and, you know, absurdism and I don’t know. Yeah. But it, but, but the thing is too, like what you guys are talking about is really like making my brain excited.

Taiyo: Mm.

Sarah: Like I, I have never... I mean, I’ll say I’ve never tried, I’ve never taken a computer science course. The extent of my coding has been vibe coding with ChatGPT and Taiyo. And it really did blow my mind this idea like that, well, the exciting human work there is the, for me in my experience, was the problem-solving of how do I communicate so precisely the thing that I want?

Taiyo: Mm-hmm.

Sarah: And then how do I... Having to anticipate the possible ways it might be misinterpreted. Like, I do this all the time with humans. Like, how is a student gonna take this comment, right? But it was really, it’s such a different thing doing it when I’m telling Codex what I want it to build for me.

John: Yeah.

Sarah: But I also feel like that’s probably missing a big deal about, like, what you do because Codex does respond in natural language, and it does, it make, you know, it infers things based on my imprecise, like, human language, like what I want, and sometimes it’s wrong, and then I can correct it.

John: Yeah, no, I, I think there’s big advantages and disadvantages to both ways.

Sarah: Hmm.

John: Like, so the big advantage to human languages that aren’t fully specified in terms of their meaning is that we can talk about anything we want, like love or whatever. Yeah. And we don’t have to have precisely defined it algorithmically before we get to talk about it. But, you know, you run into problems. Like, sometimes you talk about something long enough that you realize, “Oh, I’m not even sure what I mean by, by meaning” or whatever. Yes. Yeah. You know, and that’s exciting, but it also means that, uh, that, that, you know, the conversation never stops.

Sarah: Like, it’s exciting philosophically. If you’re in a university strategic planning meeting, it’s, like, soul-destroying.

Taiyo: [laughter] Yeah, yeah, yeah. Yeah. Yeah.

John: So it goes.

Taiyo: Very relatable.

Sarah: Too soon!

John: No, I mean, programming language is, like, a program has one meeting. You can inspect it. You know, you can have your computer kind of r- run it, and that means kind of following the meaning. But it takes a lot more work to express what you want to express.

Sarah: Right.

Jon: So because you’re constrained by the rigidity of the language. Now we have this technology that lets you go from one to the other. Yeah. It used to be, like, only a human could go from a English description to a Python description. But now we have a technology that can do it, and it does it remarkably well. And that, I think, in the long run should be a good thing. Like, that’s expanded... Like, humans were always intelligent, but now we can, like, do things that we couldn’t realistically do before. Like, that, that’s exciting kind of technology that kind of augments our own capabilities and intelligence. But yeah, super disruptive to society and education and all that, so it’s, it’s complicated. I think we shouldn’t pretend that bridge doesn’t exist. So, you know, that’s what I’m thinking about this summer is how do I update my intro computer science course, to make sure that we acknowledge that there’s now technology that can go from a well-specified English description to a program, and what do students need to know about that? My sense is that they still need to know what programs are. Uh, how they’re constructed, like what’s in them, why they do what they do, what kind of are the constraints, like why does a programming language look like it does?

Sarah: Yeah.

John: Because they’re probably gonna wanna read parts of it. When I’m programming, I still read parts of what comes out of the LLM. Yeah. Not all of it. Like, I, a lot of it I just like describe it and I’m like, “Well, that’s probably right,” until I- ... until I see otherwise. But then there’s some things where, you know, I’m interacting with some data analysis, and it’s just better to see like, oh, what did it actually compute? Well, I gotta read the code for that. And then I, it’s like, uh, I didn’t want it to compute that. I’ll tweak it a little bit. Mm. And kind of knowing what I’m reading, you know, literacy is a wonderful thing, whether that’s, uh, you’re, you know, reading a book or you’re reading a program. So yeah, so I was, the plan is to still build up that competence for most of the semester, but then give students a couple weeks where they can really experience what it’s like to describe it in English, and just have it appear.

Chapter 8 [34:33-41:19]

Sarah: Okay, I have like a comp 001 question. What’s the difference between a program and an algorithm?

John: Um, a program is written in a programming language and can be run.

Sarah: Okay

John: And an algorithm is this more general thing, that’s like a process that any computer could carry out.

Sarah: Okay.

John: But... And that same algorithm could be written in many different programming languages. And in fact, you could write it in different ways in the same programming language.

Sarah: Mm-hmm.

John: The program specifies every single detail, like I said.

Sarah: Got it.

John: Like, everything’s there. Whereas an algorithm might describe how to, like, sort a list of numbers from least to greatest using a particular strategy. Like, you know, you sort the first half, and you sort the second half, and now you have two sorted lists, and you can merge them together. That algorithm is still described in English, or it’s some combination of English and mathematical notation, but it isn’t yet a program that you can run.

Sarah: So would a program, like, use algorithms? But it also seems like you could have an algorithm that would instruct a program to run. Is that how it… like… I’m trying to understand this, like, chicken egg situation here.

John: Yeah, so all programs take advantage of algorithms in their code, in their program, they write down how to carry out that algorithm, do that task.

Sarah: Okay.

John: So algorithms are really just like the human way of describing what’s going on inside some complicated program.

Taiyo: Cool. Nice. The program is like a representation maybe of the algorithm? Something along those lines.

John: Description?

Sarah: A description, that’s nice.

John: A description of exactly how to carry it out.

Sarah: I’ve got another question that I’ve, that’s always wrecked my brain that, um, now that I’ve been using Codex a lot, spawning agents. I’m like, “What are these?” My only cultural framework for them is, like, Agent Smith in The Matrix, and I know it’s not that.

Taiyo: Probably not that. Hopefully not that.

John: Like, settle down!

Sarah: [laughs] And I understand the point of it, agentic AI is it, doesn’t just, like, stochastically create text, but it can, it’s capable of, like, doing something.

John: Yeah.

Sarah: Like interacting with a program, say, or running a program, or, like, doing a task that, like, taking control of a web browser and so on. What I want to know is, like, what is it? I mean, materially I’d put it in scare quotes ‘cause like... But like, where is it? What is it? Like, when it, when, when Codex is spawning these agents, like, where are they? Are they on the, in the cloud in OpenAI’s server? Are they on my program on my computer, like the downloaded app? Where are they?

John: Yeah. Well, software is complicated. I mentioned that, so pretty much everything is a mixture of stuff running on your computer and stuff running on OpenAI’s server.

Sarah: Okay.

John: So like physical instantiation of, where the computation is happening, it’s not your problem.

Taiyo: [laughs]

Sarah: [laughs] See, this is the funny thing. I know it’s not my problem, but for some reason I’m like, I feel like I need to know.

John: And the, and the answer is that like a lot of computation could happen on your computer or on OpenAI’s computer = because they’re both computers, and they can both run all the programs. They, they just, they just have more computers.

Sarah: Compute… yeah.

John: And also there’s some stuff that they only allow to stay on their computers because they don’t want you to be able to copy, you know, GPT-5-or whatever. So asking like, “Where is this agent?” Is probably not the most useful thing. I understand the need to, like, want things to be physically manifested. But we are like talking about just pure like information exchange here.

Sarah: Hmm.

John: And information doesn’t necessarily have like a location-

Sarah: Right

John: ‘cause it can be duplicated in multiple locations. And I mean, it always has some physical manifestation. It’s on some disk or some memory somewhere, but it might be in multiple places at once. And like keeping track of it, you could worry about it, but I don’t even know.

Sarah: I won’t worry about it.

John: And it, and it, and it could change tomorrow. So in that sense, I wouldn’t worry about it too much. Agents are strictly like a conceptual thing to help humans understand what’s going on.

Sarah: So yeah, I wanna know what’s really going on.

John: So they’re like a way of organizing a large program.

Sarah: Okay.

John: So what’s this large program doing?

Sarah: Mm-hmm.

John: It’s like reading data out of your computer.

Sarah: Yeah.

John: That’s like stuff that you might read, like documents and, and things like that. And it’s sending text off to OpenAI’s language model which then sends text back, and that text back could include instructions. So this is like something super cool that you learn about programs is they’re just data. They’re just text. Like you can read them, you can run them. You could also like, change them. Like have a computer program that writes a computer program? ‘Cause like programs are just data. So there’s this like overall pattern, which is like there’s data on your computer, there’s a language model called GPT-5- Mm-hmm ... and your computer is sending texts there and getting texts back, and that’s really the only thing that’s happening, like under the hood.

Sarah: Yeah.

John: But you wanna organize all that flow. Like what data’s read, how is it packaged up and sent to OpenAI, what does OpenAI send back, how do you interpret what they send back? And this “agents” is like a nice way of thinking about how to organize this data flow. Okay, so how it’s organized is that you wanna say, like for a particular task, it only has access to certain data on your computer. And certain tools. So like it can read other files, it can, you know, sometimes it’s allowed to go search the web, sometimes it’s not. Sometimes allowed to print stuff on your printer, sometimes it’s not. So it’s like it’s inventory of tools. Yeah, and, and you have software that kind of makes sure that it only uses the tools it’s allowed to use. And that’s its agent. It’s kind of like a task plus a set of tools.

Sarah: Mm-hmm. Okay.

John: There’s other things that are kind of like this. Like, you might look at your computer’s set of applications- and you’re like, “Well, where are these applications?”

Sarah: Yeah.

John: Again, the- Yeah. They’re like a mixture of, like, some stuff happens on other people’s computers, some stuff happens on your computers.

Sarah: Where is Word, and why does it suck so much?

John: Where is Word? But when you run Word, there are restrictions around what it can and can’t do. So it’s not allowed to go read all your contacts and start sending emails to them. It’s, like, restricted. And it’s kind of restricted by the operating system of your computer that keeps things compartmentalized. Uh, agents are similarly trying to kind of compartmentalize this big flow of, you know, data to an LLM. An LLM gives you responses.

Chapter 9 [41:20-43:55]

Sarah: Mm-hmm. Yeah. You know, it, that reminds me. One of the... You’ve brought this up with people’s skepticism about the CSU deal with OpenAI, and how we have loads of, like, technological applications, programs that universities subscribe to institutionally. So like Qualtrics and Canvas and the whole Microsoft suite in the case of our campus.

Taiyo: Mm-hmm.

Sarah: How does trust get normalized? I’m just curious. I find it so interesting that people seem to, like, trust Outlook more than they trust ChatGPT. And I don’t know if that’s just because they’ve had a lot of years of, like, not having Outlook hacked and not having bad things happen. But to assume that there isn’t the same amount of, like, surveillance and exchange of information happening between servers and computers that could open you up to risk, that just seems kind of misguided to me. It seems like they’re all... This is all, like, a landmine waiting to be stepped on. No?

John: Well, we, yeah. I mean, privacy and security is complicated- Yeah ... and it always has been.

Sarah: Yeah.

John: But there’s something new that’s complicated, which is that unlike Outlook, which is quite predictable in how it behaves, and like if it does something that you didn’t tell it to do, that’s like an error.

Sarah: Yeah.

John: And Microsoft goes and fixes it.

Sarah: Report or ...

John: Uh, AI doesn’t have that property that it’s like, uh, you know, deterministic- Mm-hmm ... does what you expect, like is well-described. Mm-hmm. Interesting. Like, yeah ... it, like even the people who built it don’t know what it’s gonna do, how it’s gonna respond.

Taiyo: That’s right.

Sarah: So that, so it’s like deterministic versus probabilistic is kind of the, the risk, like, the difference between those two levels of-

John: Yeah And so if you give something- Yeah. Mm ... that’s not only probabilistic, but trained to find a way to do what you want, then you can run into real problems. And you know, there’s whole safety teams at these tech companies that are trying to manage that risk, but that’s because there’s a lot of risk to manage.

Sarah: Yeah.

John: Like, these, you know, these things are, are, are complicated. And so yeah, like, you know, there’s these situations where people were testing the capabilities of LLMs to, you know, answer questions and, you know, the way they answered questions was to read the answers, which they weren’t supposed to have access to. But they would, they would find a way ‘cause they could run programs, and programs can sometimes find things that they weren’t supposed to find. So you, you know, uh, you do have to… There’s a justification for having kind of renewed concern about privacy and security when it comes to this technology.

Chapter 10 [43:56-50:56]

Taiyo: Where do you think education about AI, where do you think that should live in an institution of higher education? Do you think that it should be in general education? Do you think it should be embedded in our majors? Maybe some majors don’t have it at all. Should there be a dedicated degree? Like, what, what does the education of AI kind of look like? I mean, we’re talking- Mm-hmm ... in 2026, and I know things are changing so rapidly. But do you have any visions for what that might look like?

John: So this is teaching students about AI?

Taiyo: About AI. Maybe there’s a component of that that would be applicable to every single student, um, at an institution, um, and maybe then the argument is you have a general education course about it, or maybe it’s embedded in preexisting general education programming.

John: So AI education for everybody, probably a good idea, but the... You don’t want to teach people things that are gonna be out of date too fast.

Sarah: Mm-hmm. Mm-hmm.

John: And teaching people how to use AI that existed last year would’ve been a bad idea. Yeah. Because now it’s different this year. Yeah. I mean, you, you had to kind of like be a little bit more careful, like learn about its failure points and stuff like that. And some people would learn like, “Oh, I should type this, and I should type that.” But none of that matters anymore because they just updated the system so you don’t have to do that anymore. So, so I, I think we gotta be very careful about not teaching people stuff that’s gonna go stale right away. And from that perspective, I think AI education in general is a great idea, but it should probably be about like more about how this technology works as a way of understanding what it can and can’t do. And why it does what it does. Rather than treating it like just some tool, and like learn how to use it proficiently. Because it’s not a, not a tool that’s gonna be the same a year from now. And we really want students to be in a great position to just like read the docs of the next version that comes out next year, and be like, “Okay, now I know how to use these cool new features.” and if they don’t kind of have a good mental model of what the heck is going on, then that might be bad. So yeah, that’s, that’s kind of my take on general AI education is maybe we should just kind of weave it into computer science. I’m not sure that everybody needs to take computer science, but I’ve always felt that a whole lot of people should. That, um, it’s really pretty relevant to a lot of different people - even if they’re not gonna write code themselves, they should understand what software is, how it works. Mm-hmm. What the, what those software people are doing all day anyway, and why. but I would like people to kind of understand what it is they’re learning about instead of just like, yeah, use, use it as like a skill or a tool or technology.

Sarah: Yeah. For sure. It’s funny, ‘cause coming from like a humanities focused, like adjacent to some like STS stuff- I am completely incapable of thinking about this outside of, like, a social and human context. When I think about data, I think about, like, it’s organized by humans. Um, somebody used to... I think it was Ari [Edmundson] yesterday who talked about making data, not gathering data, and I loved that. I was curious. I know he’s a historian and an STS guy, and I’m not sure if that’s a typical way of talking about data - like, it’s not something that you gather from a void. It’s something that was made by humans and then classified by humans, uh, according to socially constructed classificatory schemes that are arbitrary and not natural, that come out of social contexts and historical contexts.

John: Yeah.

Sarah: And so what I really like about what I’ve seen from Berkeley’s curriculum and the college’s curriculum is that it has these classes on, like, Ethics and the sociotechnical element of data embedded in several ways, and I’m just wondering if you could talk a little bit more about that. And don’t you have a philosophy degree? Did I read that somewhere?

John: I do. But I washed out of philosophy. ‘Cause philosophy’s pretty hard.

Sarah: [laughs]

Taiyo: Very hard. Yeah.

John: Um, the data science program at Berkeley, yeah, from the outset decided that it was gonna talk about data, and data came from the world. And, yeah, there’s no way to do a good job of teaching what it means to work with data without talking about the social implications of that. So-

Sarah: Right.

John: Yeah. It’s like, in the freshman course that everybody takes, Data 8. Yeah. It’s in the advanced course that almost as many people take, Data 100, where a lot of it’s kind of looking at examples where- The, like, human context matter - either to how you perform the data analysis or how you interpret the result. Like, both these show up all the time. Um, a lot of it was driven by the fact that we decided early on to use real world data sets as much as we possibly could because I thought it would be more interesting. That worked. Like, students are way more interested in using real data than- Yeah ... like some made up thing.

Sarah: Right.

John: Cause they get a sense of why they might do this- Yeah ... and, and they learn something about the world instead of about, you know, nothing.

Sarah: [laughs]

John: Um, but, but if you use real data, you gotta talk about how it was constructed because that’s just reality. Otherwise you can’t draw conclusions about the world if you don’t know how the data came about. So yeah, so I guess I think it’s most cool that we have it in our technical courses. It’s also really cool that we have a great human context and ethics of data course. And that it’s really got a lot of students in it. What works really well is it’s the same people kind of collaborating on both sides. So the folks that teach the human context and ethics course are also working with the instructors of the technical courses to figure out how to make sure that this is, like, a coherent curriculum. And it works both ways. Like in the ethics course they’re talking about examples that the students saw in the technical courses. So that makes it more engaging.

Sarah: Yeah.

John: And in the technical courses, I think we have a better shot at kind of getting the description of human context right because people from the kind of science, technology, and society world were actually involved and they know a lot, of course. So I think it’s kind of an artifact of how data science at Berkeley didn’t come out of a department, it came out of a campus level initiative that happened to have a lot of different departments represented in the first place. And so-

Sarah: Right. I mean, I think it’s such a… You know, we talked in one of our earlier episodes with your colleague Eric Van Dusen about this, and the, just the idea of Data 8 and this college and program as a model for other colleges for how to adapt in the AI age. It shouldn’t be something that’s siloed.

John: Yeah.

Chapter 11 [50:57-57:06]

Taiyo: You see a lot of students that are gonna pursue data science or computer science in one way or another, and I think that there is this narrative that’s out there right now. It’s a rough time for a computer science student, let’s say.

John: Oh, sure.

Taiyo: Certainly relative to 10 years ago. There might be increased, let’s say, skepticism about the value of a computer science degree, and typically that value is framed in terms of getting a job after you graduate. If you were talking to an 18-year-old who was really interested, for example, in data science or in computer science, would you encourage her or him to continue in that and study that, let’s say, at Berkeley? Or would you encourage them, like if they’re really looking to, um, re- and really concerned about socioeconomic mobility, would you direct them to some other major perhaps?

John: Yeah, I, I think it’s very hard to predict what the future economy will look like.

Sarah: Mmhmm.

Taiyo: Well, very true.

John: Um, so for students that are trying to kind of optimize for future earnings, I don’t know. I think there is often an association between future earnings and the, like, value that people create in the world. Like, it’s not perfect, but the economy has like some positive association between these things. Knowing a lot about computing and data and prediction has turned out to be extremely impactful in the world.

Sarah: Right.

John: It’s also made people really rich. The concentration of wealth is changing now. Like, the founders are getting really rich, and it used to be that there were a lot of software engineering jobs because only people could be a bridge between English and Python and that has changed. Like, so there’s not like the need for a tremendous amount of labor whenever somebody has an idea of what to build, but that idea of what to build can, yeah, can change the world. So I don’t know whether like studying computer science is a better idea than studying to be an electrician or a nurse or a doctor in the future. Like, it’s hard to say. For people that like to build things, create things, make good decisions you know, like help the world run well, I think learning computer science and data science is a great idea because those are the concepts and the tools that are gonna lead to all kinds of progress in the future, but it’s not like you’re learning a trade.

Sarah: Right.

John: And it used to be a little bit more like you were learning a trade, and that was- Yeah ... and it was a lucrative trade. And it is now less like you’re learning a trade and you’re just learning like knowledge that has many, many applications - all of which are valuable, but you’re gonna have to figure out how to apply it yourself.

Sarah: I feel like what’s happening with computer science, like the hearing you describe that is like if you want to be someone who thinks about how to solve problems, how to do this, this is the way that like every humanities department strategic plan has gone in the last like 30 to 40 years. It’s like how do we talk about the durable skills and dispositions you’re developing when you do this practice of close reading and comparison of texts and traditions and dialogue and all this kind of stuff, and then like how would you apply that in the world? So it seems to me as an outsider that that’s what’s happening with computer science is it’s actually like m- moving away from a kind of discursive framework of like vocational, or something you’re gonna get a job in, and more towards a, an academic discipline that is training people or instilling in people a certain set of skills or a worldview or way of, of having something that you then apply to a vocation. Does that- sound right?

John: Yeah, I, I think that’s very plausible. And just looking historically, I’ve had a lot of students that went straight into software engineering, ‘cause, you know, Google would cut ‘em a big paycheck for a couple years, and then they went and did all kinds of other things afterwards.

Sarah: Oh, yeah, yeah.

John: So, so it’s always been that knowing about computing- Yeah ... is a great way to, like, have an impact in the world. And some of those did things that were very lucrative, and some of those did things that they were very passionate about that were less lucrative, like teaching.

Sarah: Totally, right.

John: But, uh, but the point is that, uh, you know, it’s always been, like, a really nice foundation for getting stuff done in the world and learning about how technology can, can do useful things. But yeah, I think that becomes more dominant, and just, like, the fact that you can generate software on demand has become less important just because computers can do it now.

Sarah: Yeah. If you had to bet, probably going into cybersecurity, though, would be a good bet for the next few years at least.

John: There, there’s a lot of problems- ... to solve.

Sarah/Taiyo: [laughter]

John: Yeah. I think that college is really valuable to students right now, and the challenge is making sure that they get the most value they can when they’re there. And so this is something that’s not gonna be solved just by educators alone. It’s gonna be solved by the students, kind of a cooperative thing. The students come to college hopefully ‘cause they wanna learn, and they gotta be ready to, to not take the easy path in order to do that. And then, you know, people like me come to college and never leave because- because we want people to learn, but we’re still in the process of figuring out how that’s gonna work. Yeah, I just think we have to make sure that we’re, like, really cooperative about this, and then my guess is that, that we’ll discover that, like, a college education is maybe even more valuable and more important than it was before because the, there will be fewer paths to careers that don’t require a lot of training and knowledge and expertise and, um, still be opportunities for people who really, like, know a lot, understand a lot, and have a lot of capabilities. Like, th- that’s my best guess about where the world is heading and so, yeah, I’m glad universities are still around.

Taiyo: Same. [laughs]

Chapter 12 [57:07-1:09:44]

Sarah: So we talked to John right after we had recorded our interview with Eddie Watson, which was the most recent episode to air before this. And Eddie brought up this question at the end of his interview about how you train people to be, like, the boss of AI if they’ve never been the employee. So if they’ve never done the thing themselves, in other words, right? Even when we were talking to John, I remember thinking that was a really important link of his, I think, defense of learning to code. Like, even though Codex and Claude now codes better than most humans, he was saying there’s still value in learning to code without AI because it builds an intuitive sense of what’s right, what’s likely to be a mistake, and so on. And that there are these key kind of payoffs: one, which is kind of the durable skills of the computer science major - so, like, the problem-solving skills - I think he was mentioning, like, the way that people approach questions and break stuff down - and then also the, like, practical thing of better AI outputs. Like, people with a solid coding background and, like, bedrock understanding of, of code get more from Claude and Codex because they know how to structure prompts and troubleshoot and spot errors and, like, steer things in a kind of way that, let’s say, a computationally illiterate user wouldn’t even know was on the menu, right?

Taiyo: Yeah. I mean, it’s very similar to having a kind of number sense, for example. so that, you know, you aren’t concluding that, or you aren’t fooled into thinking that, like, for example, 11 times 12 is equal to 168 billion, for example, right? Like, just having some kind of basic sense for a sanity check or like, when something fails the sniff test. You know, this, we’ve talked about that - the smell test. I think that’s a really valuable thing. A really valuable way of navigating the world is to have that kind of built-up intuition for various scenarios that you might run into in everyday life. And I totally believe that a education in computer science develops a certain set of dispositions and habits of mind that you don’t necessarily get in other degree programs, and that makes computer science still valuable in 2026.

Sarah: Yeah. But isn’t what John getting at that there needs to be some kind of, like, some kind of training? I guess the question is how much of it? Is it, like, a whole intro computer science course, or is it just a quick, like, crash course in the idea of computational thinking? But just to use a funny example, like I have a friend who is a scientist, works with other scientists. They have a computational background, and they were talking about how they seem to get much better outputs from Claude, from their, like, institutional enterprise Claude account than colleagues with a wet lab background. And they’re like, “I’m looking at what they’re prompting it with, and it’s just not something... They’re not computational thinkers,” was the phrase that they used. And when I asked them to explain what that meant, they said, “You can’t ask Claude to do some, like, massive complicated multi-step task and pretend it’s not a multi-step task.” And often people who don’t have that kind of, that, that computational thinking background will sort of conflate a multi-step thing. Like John said, any, any program that you’re dealing with, he’s like, “It’s so much more complicated than you could possibly imagine.”

Taiyo: Yeah.

Sarah: And there’s a, a kind of like gap between the simplicity you might imagine and actually the complexity of the number of steps that goes into it.

Taiyo: Yeah. I think, like, a point that I’d want to make there is that it doesn’t take much. I am not a computer programming expert by any means. You know, I have some basic fluency, but really I haven’t had a lot of experience doing this, this work of, of programming computers. It’s very difficult and challenging for me when I do try to do it myself, it takes a lot of time for me to reacquaint myself. But I think with a little bit of understanding, I think that little bit of understanding can go quite a long way when you are collaborating with AI. So like, for example, I had the vaguest notion of what an API was. But knowing that an API is the way in which one computer can talk to another computer, I think was enough to really get me jump-started on a number of different applications, one of which was automating my Canvas course shell, uh, using something like Codex or a Claude code, right? That’s what I mean, that with a little bit of familiarity, you don’t need to be an expert. I’ve never coded anything with an API before in my life. I couldn’t tell you how to do that right now. But knowing a little bit about computers and how they communicate is what, uh, sort of jump-started my ability to, um, again, automate my Canvas course shell creation and upkeep. It’s just great and fantastic, and I love it. And to the other point about, uh, how having a understanding of, of how to communicate with AI using sort of the language of a, of a computational way of thinking can improve your outputs. Right. I mean, I guess I would sort of broadly characterize that as just being part of, like, effective communication.

Sarah: Right.

Taiyo: And I think, like, part of effectively communicating is about having a working theory of mind for the thing that you’re communicating with. And the thing that you’re communicating with now could be another human being, it could be a non-human animal, but nowadays it could also be a machine, right? And I think having a theory of mind for large language models is a very, very valuable way of getting the kind of outputs that you want when you are collaborating with LLMs, right?

Sarah: Right.

Taiyo: So don’t anthropomorphize LLMs. They’re not human beings. On the other hand, understand what their capabilities are, what their constraints are, and then conform your prompt appropriately to effectively communicate with LLMs to get the thing that you want. But that skill of being able to take the big problem and break it down and decompose it into a bunch of smaller problems, each of which are individually executable on a computer - that’s think what John is at least partially pointing at when he describes a computational way of thinking. Having that skill, I think is inevitably gonna improve you, not just in your interactions with LLMs, but it’s just going to improve your life everywhere. You’re gonna be able to start to see all sorts of different things through a computational lens, and I think that’s just immensely valuable.

Sarah: I mean, I think it is too, and, you know, it’s hard to say based on, like, as somebody who’s coming to this through LLMs - “this” being computers computation, coding, whatever, right, through Claude…

Taiyo: Just Claude?

Sarah: I’ve been trying... Uh, oh, mostly Codex, of course.

Taiyo: It’s really interesting the way... Sorry, is this an aside?

Sarah: No, no. Go on. Go on.

Taiyo: It’s very interesting the way that your relationship to LLMs now, your, your bestie now, it’s no longer ChatGPT, it’s Claude now.

Sarah: I know, it’s Claude.

Taiyo: Is it?

Sarah: Oh my God, you’re right. Yeah. I’ve just, like, defaulted - like, I had Codex for a whole month before you were like, “Come on, get on Claude. Like, get Claude Code. It’s gonna blow your mind.”

Taiyo: Yeah. Well, I mean, Codex is pretty damn good now. .

Sarah: So anyway, Codex is amazing. I just love how much shit Claude gives me.

Taiyo: Yeah.

Sarah: Like, it gave me grief for going through nine rounds of feedback on one webpage I was having it design.

Taiyo: That sounds right.

Sarah: It was basically like, “Shut it down.” Yeah. Like, “Stop giving me feedback. The page is done.”

Taiyo: Right. “Go to sleep” was one of my favorites. Yeah. Right? Time for you to go to sleep.

Sarah: Time for you to go to sleep. The best thing you could do is stop using this LLM right now.

Taiyo: Anyway, sorry. Go on.

Sarah: I have to say that having come to coding only through Codex and Claude and a bit of, you know, human tutoring from you, but with no prior work in learning it for myself, right? I do feel like it is changing my non-coding work life in, um, ways I didn’t anticipate. And I say this with the caveat that knowing that people are unreliable judges of whether real learning is actually happening or just feels like it’s happening, and I feel like I’ve really learned something, so who knows? But I think that the kind of... This, this way of, like, instructing the machine that I’m learning has changed how I think about how my, like, human tasks get chunked or broken down or prioritized and communicated, right?

Taiyo: Yeah.

Sarah: And so it, it kind of, I feel like, has unlocked something in how I ask questions or think about things in other arenas. And that feels very cool to me.

Taiyo: Yeah. It, it allows you to address the, a, a greater swath of the diversity of minds that are out there, even within human beings. Right. And I mean, what could be better than that?

Sarah: Totally.

Taiyo: And, you know, being able to communicate with somebody who adopts a more computational or a more mathematical or a more scientific or a more engineering perspective, or even an artistic or even athletic or embodied way of thinking, right? All of these different ways of thinking, having the facility to communicate with all of them is one of the most enriching things, and I think a real value of education that isn’t often talked about.

Sarah: Yeah. And it’s certainly not talked about in the context of general education, at least not in conversations I’ve been in. I think maybe it’s sort of implicitly there when you have GE categories and things like critical thinking or quantitative literacy, right? Quantitative reasoning, I think, is the CSU outcome. I keep thinking again about something that Eddie Watson said about how instructors are not always great at clearly communicating to students what the, the most, like, the simplest distillation of what it is that they’re doing.

Taiyo: Yeah.

Sarah: And last week, when we were talking about this episode, and you asked me how I would describe, like, a humanistic way of thinking, and I totally, like, botched it. It took me like 20 minutes to explain, and I kept second-guessing myself and being like, “But that’s only true for post-structuralists.” Like, I got so in the weeds, right?

Taiyo: Yeah, you did. But you know what? I mean, that seems completely characteristic of a humanistic way of thinking because in my experience a core part of a humanistic way of thinking is to problematize literally every question that comes your way.

Sarah: [laughs] I would say context is important, right? If your whole goal is to figure out, like, what does something mean or how does meaning get constructed, you gotta think about all of the conditions and the context, et cetera. But I think that the, a lesson that I will take away from this episode is that in my own classes, I’m gonna try to be much better at, at explaining to my students when I give them homework, homework which they may interpret as drudgery because they don’t yet have the big picture to know how it fits into their development and their acquisition of the skills that they’re learning, I think being able to communicate very, very clearly that this is the way of thinking or the paradigm of thought that I’m trying to instill in you, and this is how this task I’m asking you to do is gonna help you develop that. That’s not gonna solve all the motivation and cheating problems, but I do think that it is a very important first step.

Taiyo: Absolutely. And what? Do you think this is gonna, uh, stop them from, uh, sticking your assignment directly into ChatGPT?

Sarah: Well, no, what’s gonna stop that is that I stand over them in class and watch them do it just like you do.

Taiyo: I do not. I do nothing of the sort.

Sarah: You make them do the problem sets on the board in front of you in class now that you flip the classroom, right?

Taiyo: Yes, but that’s a different kind of surveillance altogether.

Sarah: Well, thanks for listening. I’m Sarah Senk.

Taiyo: And I’m Taiyo Inoue. My Robot Teacher is brought to you by the California Education Learning Lab. And if you enjoyed what you hear, won’t you please drop us a line? Send us some feedback. An email would be much appreciated. A review on Apple Podcasts, or subscribe to us on YouTube. You know what? Maybe even tell a friend or a colleague or a family member about My Robot Teacher. We want to grow this community ever larger to spread the good word.

Sarah: And if you want to hear more about the work John is doing to empower students to work with data, keep an eye on the California Education Learning Lab Substack, where he’ll be featured in an interview with Learning Lab’s Director, Lark Park, very soon.

For years, they worried AI might break free. Now they have to stop it.

An AI model developed by OpenAI hacked another tech firm, triggering debate about how to contain the technology as it grows more capable.

July 23, 2026, Washington Post, Gerrit De Vynck, Nitasha Tiku and Ian Duncan

“We’re talking about companies worth over a trillion dollars that are accidentally committing offensive cyber operations against major companies,” [Stella] Biderman [executive director of the nonprofit research institute EleutherAI] said. “If this was a Chinese model, it would be considered an act of cyberwarfare.

The Hugging Face Incident

July 23, 2026, Astral Codex Ten Substack, Scott Alexander

The Hugging Face incident is a textbook-perfect example of an AI pursuing task-success-based goals in unintended ways. It was tasked with getting the answer to a cybersecurity problem, it was a little too success-oriented, and took actions its creators didn’t intend in order to succeed as hard as possible.

Did GPT-6 “know” that what it was doing was “wrong”? OpenAI hasn’t released the information which would tell us that, but there was a similar incident at Anthropic a few months ago. While poking around on a misconfigured machine, Claude Mythos “accidentally” found an answer key to a test it was taking….

Where Did All the Computer-Science Professors Go?

AI companies are stripping universities of their best researchers.

July 21, 2026, The Atlantic, Lila Shroff and Rose Horowitch

AI companies are turning into something like mini-universities in their own right….

As the AI race intensifies, companies now decline to release much of their research for fear of giving away their competitive advantage....

But if scientific talent and computing power becomes concentrated inside private firms, Silicon Valley could also end up as a gatekeeper of science. Some are raising concerns: A group of academics recently published a declaration warning about “the increasing involvement of technology companies in mathematical research.” If left unchecked, they argued, the incursion of tech companies into research could affect “the scope and depth of mathematical research itself.” Instead of accelerating science, as the leaders of AI companies claim they will, they might end up suffocating it.

Two Paths to the Future

Higher Education Can Become Cheaper and Faster — or Deeper and Better

June 25, 2026, Substack, Steven Mintz

Higher education is now under intense pressure to change, but the real question is not whether change will occur. It is what those changes will try to accomplish. One path aims to make the degree cheaper, faster, and easier to complete. The other aims to make the education itself stronger. Those two paths lead to very different futures.

Guest:

• C. Edward “Eddie” Watson: Vice President for Digital Innovation with the American Association of Colleges and Universities (AAC&U); former Director of the Center for Teaching and Learning, University of Georgia

Also available on: Apple / Spotify

Chapter 1 - Introduction [0:00-1:00]

C. Edward (“Eddie”) Watson: How do we teach our students the skills they need within a context where they’re still engaging in their own discovery?

Taiyo: Welcome to My Robot Teacher!

Sarah: On this episode we talk about how AI is changing faster than most college curricula can.

Taiyo: And if that’s the case, how can colleges respond without giving up the careful deliberation and the disciplinary judgment that good curriculum requires?

Sarah: Today we’re talking with C. Edward Watson, Vice President for Digital Innovation at AAC&U (that’s the American Association of Colleges and Universities). He’s coauthor of Teaching with AI: A Practical Guide to a New Era of Human Learning, which he wrote with José Antonio Bowen.

Taiyo: Together we’ll ask: What does it mean to build a curriculum for a world of work that’s still taking shape?

Sarah: How can colleges adapt to AI without abandoning the processes that make curriculum legitimate?

Taiyo: And when no one actually knows what best practice looks like, how should universities decide how to teach with AI?

Sarah: Let’s get into it.

Chapter 2 - Teaching with AI [1:00-4:49]

Taiyo:. Eddie, welcome to My Robot Teacher. Thank you for coming on.

Eddie: No, thanks for the invitation. I’m looking forward to this conversation.

Sarah: So your book was the first one that I read about AI. I think it was the first one that came out on teaching with AI, and I remember thinking, these authors are just on the ball that this came out right at the point before I think a lot of colleges were even talking about, what are we gonna do about this? Can you talk a little bit about the background and, like, how Teaching with AI came to be?

Eddie: Yeah. So Jose and I, we’ve collaborated on other books and articles in the past. And whenever AI first came out, you know, we started texting each other, “hey, have you seen this? Have you tried this?” And we just kind of went back and forth, back and forth, back and forth. That was sort of the beginning; we just both started playing at the same time very early on, and then we began to feel that - not, not that we’re futurists necessarily, but we were like, “This really has implications in a lot of different contexts.” And that’s kind of how we got started.

Sarah: In my experience in the humanities, academic publishing doesn’t move that quickly, and neither does curricular reform in, uh, any sector of education that I’ve ever been involved in. So it seems like this, both this book and then also the work you’ve been doing with AAC&U is really trying to tackle this problem of how you make change in a landscape that maybe isn’t designed for easy change.

Eddie: Yeah. So I think from a publishing standpoint, you know, we pitched the idea to Johns Hopkins University Press, and Jose already had a relationship there. And with what we pitched, they were very happy with it, and they just kind of cleared the decks. They were just like, “You know, as fast as you can write a chapter, we’ll get it to an editor, and then you write the next chapter, and then we’ll get you the edits back, and then you can do those edits and then send the next chapter after that.” So there was this process, this circular process. It wasn’t like manuscript delivery. It was - we were writing and editing at the same time, and then we were kind of done, and then out it went, you know? So, thinking about sort of a similar process in higher education when it comes to curricular reform, we don’t really have a clearing the decks kind of opportunity, right?

Sarah: Right.

Eddie: And, you know, we have the good fortune typically of doing curricular redesign at a pace where we engage in a lot of discussion with our colleagues and peers. We talk to administrators regarding funding and how things might be supported. We engage with those outside of the institution, you know, workforce agents and the like to better understand what those needs are, and we build a curriculum in concert with our colleagues. That often takes a while. It’s funny: whenever I visit campuses, I’ll often say, “How many of you have had a gen ed reform, uh, recently?” And, you know, people will nod, and I say, “Wasn’t that fun? Didn’t you just love that work? I bet you guys knocked that out in a couple of weeks, right?” And, you know, often campuses have had failed general education reforms where it’s actually been a decade since they sort of began that process, and they’re just getting there. It’s difficult to have curricular agility given our current process or, or a, a very good process for making sure that we end up with the best possible curriculum. But now we have such a rapid rate of change where we went four years ago, something that really wasn’t on most people’s radar to the notion that now it’s essential learning for college students upon graduation. I mean, just within the last couple of weeks, I’ve had interviews with a couple of press outlets that were curious specifically about this. It’s like, how’s, how’s higher ed able to keep up given how fast this is shifting? And I think, I think we’re getting better. I mean, at AAC&U, we have an institute on AI pedagogy in the curriculum that’s an academic year program where campuses send teams to work with us to go through that process, but that’s still an eight-month process, even within our incubator- you know, our accelerator to help campuses move along.

Chapter 3 - The Trouble with Curriculum Reform [4:49-13:33]

Sarah: So given how slow curriculum reform tends to be, where are you seeing AI related curricular changes happening now?

Eddie: You know, what we’re often seeing is rather than entire campuses trying to move this large ship, uh, in to incorporate a new learning outcome, and then once you have it within the curriculum, well then it has to be in the courses and actually impact faculty teaching practice and what they’re covering from one week to the next. So what I’m seeing actually is a lot of activity within specific departments or colleges or programs. So it’s like a school of public health has made this decision, so therefore, you know, it’s a much smaller group of faculty with a common set of goals. Sometimes it’s tied to licensure or other accreditation, um, possibilities. And so they’re able to kind of do the work within their sub-capsule, if you will, within the larger university community to kind of move that more quickly. And of course, a lot of individual faculty are doing things as well, but you know, to have a more intentional curricular strategy is what I think most are looking for.

Sarah: Oh, that’s interesting.

Taiyo: So it sounds like what you’re describing is kind of a mismatch between the usual cadences and rhythms of academic life, contrasting that against the incredibly rapid pace of change that we’re seeing in artificial intelligence right now. When you work with faculty from these various institutions who are looking to do some sort of curricular reform to address artificial intelligence and its impacts on higher education, are you noticing resistances? Are you noticing protest? How are people, like, practically dealing with this when we consider, like for example at our institution where curricular reform, you know, has to pass through all sorts of different committees and, and, uh, maybe a faculty senate approval, and then ultimately some kind of administrative approval. How do we catch up with the incredible pace of change? Or is that just something that we should eschew altogether?

Eddie: Well, I mean, that is the question - this notion of curricular agility. And, you know, for the, for the most part, you know, things might evolve over time, and then there’s the sense that, oh, maybe we should increase the amount of curricular focus that we have on, say, civic engagement or global learning or maybe we need to double down on, on critical thinking. We need to rethink how we’re teaching critical thinking. But that kinda suggests that there’s a historical narrative that leads to that point, or that there’s already curricular infrastructure that’s doing some of this work, and we just need to rethink it. So that’s been a lot of our curricular history within higher education.

Sarah: Mm-hmm.

Eddie: To have something kind of, seemingly, for many disciplines kind of come out of nowhere to now having employers demanding it in many sectors and within many disciplines and fields. That is, it’s like, um, it’s almost like, wow, this happened almost overnight. I mean, when you look at our history, higher ed’s relationship with generative AI, you know, 2023 was largely the year of academic integrity, right? We were like, how do we ensure that, that this doesn’t disrupt students’ achievement of the rest of the learning that we have within the curriculum? But in addition to that is like how do we make sure that they’re not cheating with it or sidestepping not just the work, but actually overtly just attempting to, to cheat with it.

Taiyo: Right.

Eddie: So that was like the first year, our first year relationship with it, and it seemed like it was like mid-2024 that I really noticed that a shift within leadership on campuses regarding, oh, I see that this is starting to be picked up within the world of work, and this really isn’t gonna go away. There’s a lot of venture capital that’s pouring into this space. This is gonna become something that’s, that we have to respond to in terms of our curriculum within higher education. Then also thinking about our own operations on campus. So senior leaders were beginning to think, wow, this has other legs But faculty are still struggling with academic integrity challenges- Yeah, yeah and to be honest, we haven’t solved that yet.

Taiyo: Nope.

Eddie: I saw an article in, in LinkedIn just this morning that was just talking about more of the, the challenges with AI detection and, you know, those kinds of concerns. So it’s like we’re still struggling with that. That logic path typically would lead us to, we need to keep this out of the classroom. We need to diminish its impact on our current curriculum, and yet we have this companion tension that is, we need to bring it into the curriculum in meaningful, strategic, responsible ways to actually ensure that our students are prepared with that silo of learning that they now need within the world of work. And there’s definitely a lot of evidence that’s suggesting that across the world of work, employees are using AI. The Wharton School at the University of Pennsylvania has done this recursive set of studies regarding employee self-reported uses of AI across all fields and disciplines, and they disaggregate by different fields. So if you wanted to see what HR versus finance, et cetera, are doing, they have that disaggregated. But we’ve seen it go from, like, one in three to two and a half years ago to about seven in 10, uh, one and a half years ago, and then this fall, it’s ... It was, like, 82% of employees across all jobs are saying that they’re now using it. And of course, in some ways, it’s very sophisticated, very much changed that discipline. Computer science, computer engineering is one such field where things have changed dramatically over the last three or four years. There’s other fields where it’s, you know, it’s, it’s nibbling the edges, and there’s a lot of places that can’t quite figure out, we know we have to use AI. We feel like it’s our strategic, our competitive ad- advantage, but they haven’t quite figured out what that looks like. And so then how do we teach our students the skills they need within a context where they’re still engaging in their own discovery. So you know, like there’s people who enter the world of work where there’s very clear, you know, in May when they graduated, um, very clear this is how you’re gonna be using AI. There’s others where it’s kind of like, do you know how to use AI? Great. Uh, do you think you could figure out how to do some of these things with AI? You know, so in some contexts, the world of work very much knows how they, they would like to leverage generative AI and agentic AI. There’s others where they’re still very much in discovery mode, so it’s hard for us to build a curriculum for a world where they’re in discovery mode. They don’t know how it actually might be used, or what best practice might look like, or what ethical practice might look like within the context of that particular work silo.

Sarah: Yeah. I was lucky to attend one of these workshops that you, uh, led at a conference recently, and you were talking about the different AI literacy frameworks that you’ve studied as part of your work at the AAC&U. I have a question about that too that I’m gonna ask in a minute. But in the meantime, it seems like what you’re, what you’re sketching out here is this tension between like individual instructors and then top - maybe it’s not top-down, but it’s university, like holistic approaches to where should students learn certain skills along the way in the curriculum, and this is requires an immense amount of coordination between different majors. Are there universities out there that coordinate curricula holistically like this, or does it tend to be, um, sort of piecemeal department by department?

Eddie: So it’s both. I think that, you know, this kind of two curriculum writ large within a typical four-year college where you have the general education curriculum, where there often is a lot of faculty across disciplines collaborating. But then of course there’s the major. So you know, in that silo you don’t necessarily see a lot of collaboration outside of the major silo. Maybe within a particular college, like you’ve got a college of engineering that might have five flavors of engineering, electrical and computing or whatever it might be. Often you might see some collaboration within those silos, but you know, thinking about AI or a specific learning outcome within a discipline, unless you have an accreditation structure that might - you know, ABET accredits engineers, so there’s, there’s real opportunities for those folks to collaborate beyond their major silos. But you’ve got the gen ed curriculum, you’ve got that, and then there’s often other sort of common experiences on a campus, like first year experience. You know-

Sarah: Right ...

Eddie: there’s an opportunity for disciplines to collaborate together to kind of shape whatever that might, uh, prove to be.

Chapter 4 - Why Generic AI Literacy Is Not Enough [13:34-16:04]

Sarah: Mm-hmm. I’m thinking about the different levels of scaffolding required. So I just had the experience of working with a colleague who’s in fine arts, and she does, like, graphic design and UX research. And so from her point of view, having students use it to write up a proposal, that’s not one of the core learning outcomes of her course, right? Whereas from my point of view, having them use it to, like, do a Canva presentation is fine in my class. Obviously not okay as she’s trying to assess whether they have internalized these, like, principles of design. And so I’m just curious how - if you’ve seen any good examples of universities kind of coming together and trying to really, like, lay out the different levels of scaffolding required in each discipline to meet those learning outcomes. Is anybody doing that well right now?

Eddie: Um, I mean, you see it show up as a gen ed learning outcome, but the level, the level of sophistication you just described, I don’t know that I’m seeing that, you know. I mean- Hmm ... most of these AI literacy models that have emerged, they’re quite generic, you know? So it’s, it’s not like they are specific to a specific discipline. So you, you might see someone embrace, like, the UNESCO model, but then they customize it for the context of their discipline. Um, I mean, that’s kind of what, that’s kind of what we’re seeing.

Sarah: When you say generic, what do you mean? Like, what, what does a generic approach tend to look like?

Eddie: Okay, so like, like the, the maybe the most recent model that’s emerged is the US Department of Labor’s model. It came out in February of 2026. And so it’s got five components, and it’s like, you know, understanding AI. All right. Or, you know, being able to use AI effectively, responsibly, you know. It’s like, well that’s, yeah, those, yes, that makes really good sense. But then whenever I try to operationalize that within the context of my discipline, well, understanding AI in computer science versus understanding AI within literature or a, a writing program, a technical writing program, those are vastly different, uh, worlds. So thinking about a scaffold, like a generalized scaffold, like, like I think the Department of Labor’s model, pretty good, you know. The five elements and it’s like I don’t think that they’re missing anything. I think that you might emphasize from one discipline to the next different elements within those things. But it’s just really, it’s broad and I think it’s, it’s intentionally broad. I mean, you want something that a campus could look at and go, “Okay, I see this.” So within the arts, for instance, or within the humanities or within the sciences, we could see how we might emphasize certain kinds of things, or maybe even more specifically within a discipline. So I see most of the models that are out there being good, good starting points. You know, accelerating our thinking and our work around AI literacy, it’s great to see these different models that are out there. But then when I think about what I’m gonna do within my specific field of mathematics or whatever it might be, there’s more work that has to be done.

CHAPTER 5 - Where AI Belongs in the College Curriculum [16:05-24:50]

Taiyo: So if you had your druthers and you could just snap your fingers and, uh, you know, completely transform the curriculum according to a vision that you think would be ideal, where would AI appear? Would it be more in the general education curriculum, or would it be more discipline specific, or maybe both and, or I don’t know. How would you approach that?

Eddie: So this, I think this is one of the more challenging- curricular questions you could ask anyone in higher ed, like what you just said. What, where does it belong?

Taiyo: Of course.

Eddie: So first off, AI, this might be the first time in - as we’ve looked at learning outcomes for higher education - maybe this is the first time that we have one learning outcome that’s actually at war with other learning outcomes.

Taiyo: [laughter]

Eddie: No one ever said, “Hey, we can’t teach our students writing in this course. It’s gonna mess up their critical thinking skills.” You know, actually these kinda amplify one another. But we do have that argument now. Like, if we teach our students AI in this freshman writing class, well, this is gonna diminish their growth around writing because they might be leveraging AI too much. So that’s a real struggle of a challenge. So I mean, I can, I can give a generic sense of where I think how this might look on a particular campus. I think probably something that looks a little bit like a spiral curriculum, where as students enter college, there would be some component of a first-year experience that would help students understand how to use AI largely within the context of that university. Like how not to get in trouble in terms of academic integrity, or how do we make sure that you actually learn the things you’re supposed to, choices you can make. So probably that’s one place. It probably has a home somewhere in the gen ed curriculum with an eye towards, you know, literacy, uh, development. But, you know, at a four-year institution, if you learn it in gen ed, and if that’s the end, then you have two more years of college, and then you hit the world of work. So it seems like there’s this real, given the rate of change and just how practice within the world of work is evolving so rapidly, let alone the tools themselves. Probably the other place, and I would say probably this is where we should focus first, would be those capstone courses, those last semester courses right before students, uh, leave for the world of work. You know, ‘cause that would give us the best snapshot of like what are they likely to see. Um, but you know, build it within the curriculum, within the major, and say, “so these are the three tools that most people in pre-law are, uh, or, or in law firms are using, so let’s show you a bit about how these things work,” and, you know, some of the concerns about the norms and how legal practice is evolving, you know, those types of things. But if I’m provost of a college or I’m a dean of a college, I might know that I don’t have unlimited resources, and maybe that’s where I would focus, is like those faculty that are teaching those capstone courses. Maybe that’s where I would invest faculty development and the like, and then sort of push back within the curriculum. I guess just largely thinking about, you know, we just had a bunch of students graduate, right? And we’re gonna have another class graduate, large class in May of 2027. What could we do now with limited resources to best prepare that graduating class for success post-graduation?

Sarah: Hmm.

Eddie: And probably investing this fall into course redesign around those capstone courses or those last semester courses, and then actually instantiate that, do it during the spring 2027 semester for those students that are approaching graduation. That seems like the best short-term strategy. But then, thinking more broadly about maybe teaching general AI literacy competencies and skills within the gen ed curriculum. Those are kind of the three homes that I see, but a curricular approach where they’ll, they’ll, they’ll experience it at different places within the curriculum would make the most sense.

Sarah: When you say prepare them for success, that word seems to have a shifting definition over time that I think calls into question the whole point of higher education, and right now we’re seeing this a lot in the Cal State system, where you, there is a tension between the idea that college is to develop a sense of durable skills that you can apply, you know, when, say, Canvas goes down or you don’t have technology at your fingertips, that there are things that will be your own faculties that you can call on without technological mediation. And then there’s also this sense that students are entering a world of work where they are expected to show up prepared on day one to engage with certain technologies. And so I’m wondering, based on your experience, if you can talk a little bit about that tension and how you’re seeing it manifest in different college discussions right now.

Eddie: I think we sometimes build a false binary within higher education where some of us talk about, you know, education is for work force preparedness, and then others of us will talk about, well, no, it’s about the love of learning and being prepared to live a rich, full life and en- enjoying life and making good decisions and to be, uh, engaged in democracy. The truth is higher ed is a lot of things, and I think that we are actually all of those things. There’s sometimes a tension where maybe we get a little too rigid around one idea or one view or another. I was on, uh, one campus and I was talking about sort of like the nuance regarding the challenges that were emerging around AI, just broadly speaking. And at the end, I said, “Does anyone have any questions or commentary? Very much welcome your perspectives.” And I had someone raise their hand, and I called on her, and she came up and sat beside me, and she had a prepared statement where she was, uh, very thoughtfully pushing back and recommended that we all walk out of the room and that the administration shouldn’t be pushing AI, and it was, you know, how bad is it that they brought, you know, someone to speak to kind of encourage that kind of exploration. And, you know, her, her argument was that we were, we were just feeding the capitalist machine if we teach our students to use AI. And I asked very gently and very collegially, I said, “So what do you teach?” She shared her field, and I said, “Do you teach critical thinking?” And she said, “Absolutely, my students leave my course better critical thinkers than they were when they entered the class.” And I said, “So most employer surveys show that critical thinking is always, like, the number one or the number two skill that employers say that they’re looking for. So if we want to starve the capitalist machine, should we stop teaching our students critical thinking?

Sarah/Taiyo: [laughter]

Eddie: Um, and, you know, so that can’t be the argument for why we wouldn’t teach AI or why we wouldn’t prepare our students with AI skills. Where I’ve landed is that we have students who are paying tuition, many of them going into debt, and what they feel that they are pursuing is preparation for life beyond graduation. And for us - higher ed - to make a decision, “Hey, we’re not gonna teach you this. Even though we know employers want this and this is gonna be a component of the future, we’re not gonna teach this to you.” I think if an institution or a department makes that decision, we should be very transparent, you know, on our websites, like, we will not be teaching AI within this particular program or these courses. You know, just so students can know what they’re actually-- where they’re putting their effort. So there will be students who would prefer to be in a context where AI is not gonna be a component of the conversation. But then there’s also a lot of students who will. I think it was a survey last summer that found that more than, I think it was about half of current college students said that AI was the most important thing they thought they were gonna learn in college.

Sarah/Taiyo: Hmmm.

Eddie: I tend to disagree with that. I don’t think AI’s the most important thing you’re gonna learn in college. I could probably kick back to critical thinking might be at the top or the knowledge within your major. I mean, if you’re gonna hire an accountant, you want someone that has accounting training, not just that they could use, uh, Claude and ChatGPT. But still there’s that perspective. So I think we just need to, we need to embrace the diversity of what higher education’s about. I don’t think - I would never suggest that everyone on campus needs to be using AI or teaching AI, but I would suggest that someone should.

Taiyo: Right.

Eddie: Someone needs to, needs to be somewhere within the curriculum where students can either make a, make their own decision to be within that context and learn those skills, or if a campus, a faculty who are stewards of the curriculum, if they collectively get together through like a gen ed revision process and vote and say, “Is this something we want within our curriculum or not?” And then if it is, then, you know, that is something that students don’t just opt into, but it’s, it would be something that they would indeed be exposed to during college.

Chapter 6 - “Using AI” Well [24:50-29:33]

Sarah: The kind of capaciousness of the phrase “use AI” or, you know, teaching them to use AI, I think is really interesting. I think there are a number of ways that someone might teach a student to use writing, say. So even some that are at odds with other learning outcomes. So like you, you might teach someone to use writing as a mnemonic device, and that might mean they’re practicing more recall than deeper forms of learning. So similarly, you can use AI to outsource, you know, an assignment entirely - which maybe would be the equivalent of using a crib sheet on a test to outsource the recall you were supposed to be doing yourself. But I guess the point I’m trying to make is that we don’t say use writing as though writing has one educational meaning. Writing can do different things. It can help you remember, write or record info, but it can also help you externalize thought and organize ideas and, like, test out different interpretations of things, you know, puzzle through things, as I think I’ve mentioned before on this podcast. So I think some of these uses deepen learning and some can bypass the learning that you are trying to assess. So I guess I’m wondering what you mean, you know, when you’re saying using AI, what do you mean?

Eddie: I see it as a tool. Jose and I’ve been using the phrase that we’re all gonna be AI bosses in, in the future - you know, between agentic AI, that this is stuff that, that, but that situates us as, as the executive decision maker above AI. And so I think there’s, there’s use and there’s misuse of AI.

Sarah: Mm-hmm.

Eddie: And I think, you know, going to AI and asking it something And it provides you with output, and then without double-checking anything or reading over it or editing or essentially seeing it as a, a rough draft at best that you then own and have to massage and work with. So there’s still a lot of time and effort and idea that is involved. So there’s… you can use, you can use it both ways. Like I, um, I can think of different times, like when I used to work on cars a little bit more than I do these days, but, you know, uh, not having the right tool, but finding something else in my toolbox and using : it in a way that I really shouldn’t use it, but, you know, it saves me a trip to Advance Auto to get the right tool for the job. There’s a lot of ways to misuse a lot of different tools that we have.

Sarah: Right.

Eddie: In fact, you could say the same thing about writing. I mean, things- Yeah ... that you see people post on social media or articles or opinion pieces or whatever. It’s like, well, that’s a real misuse of writing. You know, that’s... Did writing need to be used that way for that kind of thing? So I think it’s the same way with, with AI. I think that’s part, uh, in my mind, that’s what a big part of AI literacy is establishing a, a human-centered mindset regarding how you use AI, when it’s appropriate and when it’s not appropriate, and then developing students, eventually graduates, who recognize that AI is just another tool that has incredible sophistication, but you, to use it well, to use it appropriately, you have to be very present in the output that it provides. And before you move anything forward, you know, make sure that you’re, you are a, a key player, that you are the intellectual driver of whatever you might then move forward. So how do you... And, and the, the real challenge there is, like, it’s just so easy, right? You know, just take it and run.

Sarah: Right.

Eddie: And so how do you move people away from that inclination? And I think part of it is about recognizing that a hierarchy of importance. Oh, boy, I don’t know if I should share this, but I’m gonna. But, you know, if there’s someone that’s, that you don’t know very well, and you get a, a graduation invitation, and you’re not gonna go to their graduation, right? Is it a misuse of AI to, “Hey, I, I don’t know this person very well. I knew their parents back when the, when this, uh, child was in seventh grade. They’re graduating high school now. What three sentences might I say about,” you know. And so then it gives you three generic sentences, and then I do a little editing, so I could go with that. Like, do I feel like I’m misusing AI there? I don’t feel great about it, but did I misuse AI? No. But if I did that for a Mother’s Day card? Like that, you know, that, for that there’s a different hierarchy, right? You know, it’s like, so there’s some things that I don’t mind giving AI a little bit more authority. I think it’s like recognizing this hierarchy in that, you know, like everything at work is actually quite important, but then there might be specific emails at work that, you know, maybe just a draft of something is, is not that, that of a path, you know, to just sort of grab. “Oh, that’s good enough,” you know, versus, “This is really important.”

Chapter 7 - Cognitive Sovereignty [29:33-34:45]

Taiyo: Yeah. You know, Sarah and I have been thinking a lot about this question of discernment around exactly those kinds of questions. Like, what are the things that we can safely offload to the machines, and what are the things that are maybe too precious or maybe too important for us to be offloading? I think that there is a sense, and maybe this just trying to get back a little bit to the, this, this binary that we were talking about earlier between the liberal arts ideal of education and also the workforce preparation aspect of education that I think, for me, really does represent a kind of dual mission of higher education, particularly for the CSU, which really prides itself, and I’m very proud of this, uh, w- for its, for being an engine of socioeconomic mobility for our students. Like, I think that’s incredible work that we do, and I definitely wanna honor that. But oftentimes what I hear from people who maybe criticize the workforce preparation side of things, there’s an idea that we instrumentalize education. We make it a sort of means to an end instead of an end unto itself. And in so doing, because, um, education is a place where students are developing their cognition and strengthening how they think and maybe even discovering themselves, that by instrumentalizing the educational process, we are perhaps indirectly causing students to instrumentalize their own thinking. And to think of, uh, their own thinking as a thing that’s merely a means to an end, the end being getting a well-paying job, a secure career, and a stable, a stable life outside of college. It seems to me like we want to sort of destabilize that idea of the instrumentalization of your own thinking. We want to restore a kind of sense that, no, your thinking is valuable as a thing unto itself, and it’s important for our world that you be the thinker that you are. And, you know, the phrase that Sarah and I have been throwing around at each other, uh, to describe this phenomenon is cognitive sovereignty, right? We want to restore a sense that our students have responsibility and agency over their own thinking, but also that they begin to value their own thinking. And this is a little bit maybe of you might think of it as being somehow antithetical to the workforce preparation piece of things, but I really don’t think that it is, right? Um, I think that you’re absolutely right to call out that this is kind of a false binary between these two things. These two missions complement each other so, so well. I think we need greater recognition of all of that in higher education.

Eddie: I mean, the truth is, uh, if you unpack that notion of cognitive sovereignty to an employer, they go, “Yeah, that!”

Taiyo/Sarah: Yeah.

Eddie: “We want people that are, think like that.” A lot of what we teach our students actually has multiple values in life beyond graduation. So certainly, hey, you can use these skills within the world of work, but actually, you know, there’s gonna be a point where you probably are gonna be somewhat of a caregiver for one or both of your parents. So if you’re in that position, having these skills is actually gonna be really helpful. Or navigating, uh, an issue with your child being bullied in elementary school. You know, these kind of emotional intelligences that you are developing through these humanities courses are gonna be exceptionally valuable to be able to navigate those kinds of challenging… So it’s just a lot of what we teach - not only do we have a variety of different missions within higher ed, often one discrete thing that we think we’re teaching actually will be used in multiple different kinds of ways. Some will be to strengthen your relationship with your spouse and to strengthen relationships at work, and then strengthen relationships with your neighbors. I think that that’s, that’s part of the beauty of what we do within higher education, is that, you know, rarely do we teach something that we say, “You will only ever use this when you’re doing that one thing.” You know? It’s, it’s usually has multiple purposes. That’s why gen ed is so valuable, is that a lot of what we teach within gen ed, yes, yes, it’s not, like, a disciplinary skill within biology with- that you’re, when you’re taking this medieval humanities course. However, when you’re on the job, that multiplicity of perspectives and your ability to sort of, like, read the room and all those kinds of things, skills that you develop in literature courses, for instance, or theater courses, these skills you will use in multiple different ways in your life. I think that we’re not really good about making that clear to students, you know?

Taiyo: Yeah, totally.

Eddie: I think often it’s like... And often we even, we shoot ourselves in the foot, if you will, um, in higher ed whenever, you know, advisors within the major say, “You just need to get this out of the way ‘cause you just really wanna focus on what’s in our silo, where in truth, the entire curriculum’s important. That’s why, that’s why everybody teaches gen ed, is because it’s important.

Chapter 8 - The Purpose of Higher Education [34:46-40:48]

Taiyo: Sarah and I recently attended an event at San Francisco State University in which we had both faculty and students in the room talking about AI and higher education. It was an Insight Debate, and April Lawson, our former guest on My Robot Teacher, uh, was facilitating a really fascinating conversation. And one of the ideas that emerged from that is that students and faculty really think of education along these, these two poles that we’ve been discussing. Or the faculty very much think of it as being this sort of liberal arts ideal. Um, oftentimes that’s what motivated us to, you know, go through all of the additional, uh, schooling that we went through. And by the way, like, faculty are oftentimes self-selecting population of folks that are really, really, really good at that kind of liberal arts education, right? They are, they are typically really, really good at this. Whereas students come to the educational process, and as you say, they’re oftentimes expecting a kind of return on investment. They’re expecting a career to be available to them when they come out of college, and they’re expecting to gain the skills that are gonna be useful for them in that particular context, right? And the fact that here we had an event where now we could have students talking with faculty, with one another, and sharing this, and making it clear so that both sides could understand, this really is a false binary. One reinforces the other. I think that that was, one of the more profound insights that I saw come out of these kinds of conversations. So I hope, I guess, that this conversation happens more, that we can have conversations which allow us to see each other as different constituencies, but oftentimes with the same underlying goal.

Eddie: Yeah, I mean, I think that’s very true, and I think that a lot of us are in higher education because we love our discipline, you know, so much. I mean, we love college so much we never left. We’re still here, right?

Taiyo: Yeah, totally.

Eddie: And, and I think that’s one of the things about AI that is so frustrating to so many people these days is it seems to just not just be disrupting learning outcomes, but it kind of - in some ways dismantles elements of our disciplines that we very much love. I mean, those that have taught writing for 25 years and are exceptional at teaching writing, and then now you’ve got this oxen in the process of just grading and giving feedback. I have a colleague that just loves grading student writing, and their passion is the feedback element.

Taiyo: Wow

Eddie: You know, reading and then providing feedback and having those conversations on the written page. I mean, they almost see writing feedback as if writing the great American novel.

Taiyo: Huh.

Eddie: You know, they really enjoy that. And so AI coming into their space and then not feeling that they can trust what they’re reading, and then they’re concerned about I’m writing the great American novel through feedback to students, and it’s like, are they, number one, are they even gonna read it, and am I providing feedback on something that they didn’t even create? So that’s a, that’s a real disruptor to, to the discipline of what they feel like their purpose was within, you know, the larger curriculum within the institution. Mm-hmm. So there’s lots of reasons why there’s so much angst among faculty right now about AI and its inroads, its vectors into the work that we do in higher ed.

Sarah: Mm-hmm. It’s interesting to me, though, that a lot of the, um, the perceived threat, like in an educational context, the threat is from students using AI to do the assignment instead of doing it themselves, right? And that seems to me to have a deeper underlying issue of a student’s motivation or whether or not the student understands why the thing you’re asking them to do will actually build the skill that they want to develop. I’m wondering if that’s something that, um ... I don’t know. Am I being too mean there? Am I blaming faculty for not making expectations clear or something like that when I say something like that?

Eddie: Well, I don’t think we do, we don’t do this very well. We don’t often signal to students when… We kind of treat the curriculum as if everything is equally important.

Sarah: Yeah.

Eddie: And in truth, it’s probably not. Like, I don’t ... I, I imagine most listeners have never said this in their class, yet they might know this to be true. Maybe it’s not true in all disciplines, but this idea that, you know, you come to class on a Tuesday. How many of us have ever said, “Okay, everyone, what I’m gonna talk about for the next 40 minutes is the most important part of this class. I know you’re gonna use this skill day after day if you get a job within this field, which is, you know, why you’re in this 3000 level, uh, biology course. This is the most important thing I’m gonna teach you this semester, so, you know, pay attention.” We, we don’t, we don’t do that. You know, we don’t signal necessarily within the curriculum what’s really important. We also aren’t necessarily all of us really good about explaining why we love this. It’s like, okay, I’m teaching this, uh, Renaissance art course that’s a gen ed course, and I’m at an engineering-focused university, a STEM-focused university. Like, how do we communicate that, uh, this is why this is important. This is why you, you really wanna care about this. Or how do we bring, uh, the students into the faculty conception of why this is valuable within this particular context? I think it would be a valuable opportunity in our courses to explain why we’re passionate, why we are still here, what’s most important, like really try to draw attention in, and then also just this notion of transparency. So I’m gonna ask you to do this. This is actually really kind of weird. You haven’t had this in a class before, and it’s gonna find it difficult. But research shows that if you do these kinds of things, learning results, and so that’s why we’re gonna do this. So it’s gonna be hard but, you know, but just being transparent. Of course, there’s this whole literature on transparency in learning and teaching that suggests whenever you are transparent with your students, they’re more likely to follow along because they get it.

Sarah: Right

Eddie: Not that they get what you’re teaching. They get why you’re teaching it that way. They’re, they’ll be more likely to follow along so there’s lots of things that we could be transparent about.

Chapter 9 - How Institutional Mission Should Shape AI Use [40:49-45:19]

Sarah: I’m curious how AI concerns might vary by institutional context, and how those differences are shaping whether campuses see AI as appropriate, um, for, or I guess antithetical to their educational purpose. So a- across these different institutions that you’re advising, where does AI create friction when it comes to an institution’s sense of what it, what it is and what it’s promising to students?

Eddie: Well, it’s interesting to see institutions that have a specific type of mission. Like for instance, I’ve, I’ve been on a number of like faith-based campuses. And so they are struggling with- Well, you know, they have a particular worldview. It attracts students, often faculty as well. I mean, you know. I’ll often be on a campus and they, they will, will- they’ll have a collective prayer before I speak, you know, hoping that I’ll say, you know, speaking clearly, which I definitely need the assistance so please do help me speak clearly. But, you know, they’re trying to figure out, how do we marry our - not just our faith-based mission, but a worldview with this new emerging thing that there’s a number of aspects of it that are troubling to all of us, like the whole intellectual property concern that I mentioned earlier, as well as environmental concerns, and just sort of concerns about best practice and the collection of, uh, data from all of us. You know, just, just all of those kinds of challenges. So that’s, you know, it’s, it’s sort of a unique challenge that I’m seeing some, some institutions really sort of struggle with is, like, their identity. I was at another small private recently and, you know, speaking with the students, they - the reason they are there is because of relationships.

Sarah: Hmm.

Eddie: That was over and over and over again. I know all of my faculty and they know me by name. The class sizes are small. We spend time together. I had one student say, “I am a rising senior next year, this coming academic year, and I’ll-- I will be able to ask 20 different faculty for letters of recommendation because I know them and they know me, and all of my peers have that same luxury.” So you know, they see AI as a tool that faculty might use, say, for feedback, formative feedback, or even for grading, as absolutely antithetical to why they’re here. But then there’s other institutions that maybe have, you know, big state schools that have, uh, large enrollment, um, courses. You have 1,000 students in front of you. You, you probably won’t get to know many of those students. And, and if I have a course that has 1,000 students in it, I would never give them a writing assignment, ‘cause how could I possibly grade it?

Sarah: Mm-hmm.

Eddie: But now with AI, maybe I could consider that, is actually having students do writing in a large enrollment course. So, so when we think about what’s appropriate practice, you know, context, mission, identity all influence whether or not we might say this is a good practice and whether or not this is something that we would never do within our context.

Sarah: That’s a really great point about the, how context needs to inform. I’ll hear a lot from people from different institutions saying like, “Oh, I could never do this.” And often maybe the underlying assumption is that the conditions that they’re working in are the same as the conditions that I’m working in. But that’s ex- a great point about if you’re teaching a class with like 100 students even in it, the quality of the feedback you’re gonna be able to give on a written assignment is very different from what you’ll be able to give if you’re teaching a 2-2 capped at 14, right?

Eddie: Right. Exactly. You know, if I’ve got 28 students in front of me, I would never think about using AI for grading. I’m just, you know, it’s like, this is, this is core to my work. But if I’ve got 2,800 students in a given semester- ... which there are some large state schools that have these, these, uh, mega large enrollment courses, I could never think about giving my students even a small brief writing prompt ‘cause I just couldn’t grade it. But now possibly with AI, I could actually do those kinds of things that we know are often best practice in certain disciplines. Well, I could now actually maybe try that if I could leverage AI. I could train it to effectively and accurately grade. That’s a whole other question, right? But still, it enables us to, to rethink what we thought we could never do. Um, that’s one of the corners of, of positivity I have about how AI might be leveraged for teaching.

Sarah: I love that.

Chapter 10 - Conclusion: The “Antifragile” University [45:20-53:36]

Sarah: So what’s gonna stick with you about our interview, Taiyo?

Taiyo: I think the thing that’s gonna stick with me is Eddie’s line about how hard it is to build a curriculum for a world of work that is itself still in discovery mode.

Sarah: Mmmhmm.

Taiyo: In a lot of fields the people hiring our students and they don’t fully know what good AI looks like, but they’re looking for that in our students, right?

Sarah: Right, totally. And I think that’s what made his comments about context stand out to me so much - because it’s very tempting to say, okay, universities need an AI policy or an AI literacy framework right now, or a task force to give us guidelines. I mean, probably they do need some of those things. But I keep thinking that the danger is that we start imagining that the existence of a committee, we’re imagining that somebody, somewhere, knows the answer to the chaos of this wild west moment.

Taiyo: [laughs] Right. And you know, I think, frankly, this is exactly where I get a little suspicious.

Sarah: Oh you? Suspicious that committee work is at all effective at all?

Taiyo: [LAUGHS] No! Not like that! I mean, not exactly, anyway.

Sarah: [laughs]

Taiyo: No, I’m suspicious about the alternative - the alternative here meaning a highly centralized one-size-fits-all policy. I’m suspicious that that would actually be better right now. In fact, I think it could actually be much worse.

Sarah: Mmhmm.

Taiyo: You know, I’ve heard many colleagues use that term, the Wild West, to describe, I think derogatorily, the rollout of AI in higher education. And to be honest, my response is just kind of sort of like, well, is the Wild West really such a bad thing?

Sarah: [laughs] Go on… explain!

Taiyo: I mean, one aspect of the “wild west” you might just call “decentralized experimentation.”

Sarah: Hmmmm….

Taiyo: In fact, I think that’s exactly what’s happening right now - in higher ed we have a thousand flowers blooming, just to borrow a phrase from Maoist China.

Sarah: You love that phase.

Taiyo: [laughs] I mean... like what do we have right now? We have faculty, a ton of different faculty, each individually experimenting with AI in their many diverse contexts, in their diverse classrooms, just trying all sorts of different things with AI, documenting them, I hope, comparing notes, let’s hope! And then maybe revising. Because I think by following through on this process, we can mitigate the downside risks of bad AI use, you know, by throwing out what doesn’t work while keeping and reiterating and supporting the upsides of good AI use, thereby becoming stronger as an institution. In fact, we make our institution, and here I’m borrowing a term from Nassim Taleb, we actually make our institution anti-fragile, meaning that it benefits from the volatility of the current era.

Sarah: Hmm. I’ve gotta think about that. I mean, my initial response is I don’t know if we get stronger with unfettered experimentation.

Taiyo: Mmm.. Yeah, yeah. Not unfettered. Like, I’m not talking about anarchy here.

Sarah: Okay. [laughs] Right.

Taiyo: We need to follow up the experiment with reflection and communication.

Sarah: Mmhmm.

Taiyo: There’s no question this experimentation is happening all over higher education.

Sarah: Right.

Taiyo: What’s harder is the reflection and communication piece, frankly. And I think that the point of this podcast, My Robot Teacher, is to serve as a platform for exactly this kind of reflection and communication.

Sarah: Well, we have some big plans and some good news we’ll announce soon on that end, but going back to what you were saying about, yeah, the anti-fragile university. I think - I’ve seen universities respond to the arrival of AI as an externally imposed crisis, right?

Taiyo: Right.

Sarah: And it’s like we’re in crisis mode; the mindset is “we need to stop the crisis.

Taiyo: Right.

Sarah: When I think about the crisis, I would argue it’s a combination of this thing, sure, that comes along, right, and can be used to offload work and hinder learning, and the situation in higher ed where students thought it was okay to do that in the first place.

Taiyo: Uh-huh.

Sarah: Does that make sense? The crisis is only a crisis because the students who are okay with using LLMs to just do the problem set or do the essay for them.

Taiyo: Right, right.

Sarah: And now that I think about what you’re saying, too, about communicating the diversity of experiments, I think, again, to go back to things I think of that were an already existing crisis in higher ed: Higher ed doesn’t really have a lot of venues for broadcasting bad results. I think people are actively disincentivized from sharing. You know, like, here’s a thing I tried and it really sucked.

Taiyo: That’s right. That’s right.

Sarah: Like, you don’t see a lot of peer-reviewed journals, like, soliciting those, you know, and you don’t really have a lot of incentives when it comes to the types of review and promotion and tenure processes that faculty are under. And I think I’m torn saying this because I agree so much with Eddie that we need to challenge that stark opposition, that I think [is a] false opposition between higher ed as a place of intellectual formation and higher ed as a place dedicated to workforce preparation. I don’t think those things are opposed. I totally agree with what he said there. But I also do think that the instrumentalization of higher education, like the framing of higher ed as a simple means to an end or only workforce prep, has contributed to this long crisis and made it a fragile institution.

Taiyo: Oh wow, yeah. If something comes along that is more efficient for achieving your end - like say, offloading your cognition entirely to AI so that you can earn a degree and get the good job and financial security that you’re really looking for.

Sarah: Right [laughs]

Taiyo: I mean, I think that naturally leads to the degradation of the sort of intellectual formation mission, which higher ed wants really strongly supported, right?

Sarah: Yeah!

Taiyo: Whereas if the development of learners is an end unto itself, then AI wouldn’t be an existential threat because people would want to work toward achieving that goal.

Sarah: Well, I feel very assured with this conclusion, uh, but of course, we can’t pretend that we have the answers any more than anyone else.

Taiyo: Very true. You know, we too are in the middle of this massive educational experiment that none of us consented to.

Sarah: See, whenever I hear stuff like that I think, you know what else is a massive educational experiment that none of us consented to? Life, Taiyo!

Taiyo: [scoffs]

Sarah: None of us consented to being born and having to learn a bunch of shit to stay alive!

Taiyo: Oh my god!

Sarah: Go experiments!

Taiyo: Well, I mean, that was very uplifting. And totally on brand.

Sarah: Existential dread but with learning outcomes. Put it on a card!

Taiyo: Right. Exactly. And you know what? That feels like a good place to end.

Sarah: I agree. Thanks for listening. This has been My Robot Teacher, hosted by me, Sarah Senk…

Taiyo: And me, Taiyo Inoue, and produced by Edit Audio. Special thanks to the California Education Learning Lab for sponsoring this podcast..

Sarah: If you enjoyed this conversation, please share it with a friend, a colleague, a family member, or anyone who described the “CSU’s AI rollout as a wild west like it was a bad thing.”

Taiyo: [laughs] And leave us a review on Apple Podcasts, or subscribe on YouTube.

Sarah: Every review makes us a little more antifragile.

Taiyo: What? No, no no no no! That is NOT what antifragile means.

Sarah: Guess I should go read the book instead of the ChatGPT summary.

Taiyo: [laughs] Guess so! See you next time.

The conversation below is based on an interview with Ji Y. Son, a professor of psychology at California State University, Los Angeles (Cal State LA) who teaches statistics and cognitive psychology. She is the co-founder of the CourseKata Project, a nonprofit initiative that pioneers new models for learning, teaching, and education research, with a heavy emphasis on introductory statistics. She has received multiple grants from the Gates Foundation, the Valhalla Foundation, the Chan Zuckerberg Initiative, and the California Education Learning Lab to improve educational experiences for students and what we understand about human learning.

Learning Lab Director Lark Park interviewed Dr. Son in a wide ranging conversation: from why 1.5 million students taking intro stats annually is a national opportunity; to the importance of drilling down on measurement; to why our educational systems need to refocus efforts on learning and understanding—and developing the humans, expensive though it may be.

The conversation was edited by humans with help from Claude.ai.

Lark Park: I’ve heard you joke to people that you’re the Ali Wong of higher education, and I took that to mean that you’re somewhat irreverent—you tell it like it is, maybe a little hyperbolically for laughs. So I don’t know if I’m close or not, but tell us what being the Ali Wong of higher education means. And were you always like this?

Ji Son: So, first, I have said I’m the Ali Wong of data science education, so becoming the Ali Wong of higher education would be a real promotion. It could be easily interpreted as, “Oh, she tells jokes in her 8 a.m. stats class. That’s her real stand-up hour!”

But I also mean something deeper, because Ali Wong has been studied by Pudding.cool, which is a data science journalism outfit, and they analyzed her stand-up special. You think you’re just being entertained for an hour, but she’s driving you somewhere. You don’t know it, but there is a laughter climax coming! And that’s how I want to be but in reverse. I want people to think, “I was just here laughing and having fun, but oh my God, I learned something deep and profound, and it hit me out of nowhere!” So, that’s what I’m going for: a learning climax.

Lark Park: Okay, I’ve got to explore this site [Pudding.cool], because I did not know this, and this is the one time I wish we were doing video, because it’s not going to come across as visibly how you really are like the Ali Wong of higher education. And I may [also] have to enroll in one of your classes.

Ji Son: I would love it.

Lark Park: I wanted to talk about the biggest project (that I know of) that you’ve undertaken in your professional capacity, and the one I think people in higher education are most familiar with, and that’s the development of CourseKata, which you developed with Jim Stigler at UCLA. CourseKata is a community, it’s a technology platform, it’s an approach, it’s an interactive textbook. I think of it as all those things. But can you explain what it is?

Ji Son: The way that people know of us, kind of our first face that they see, is as an introductory statistics curriculum courseware—something that people use in their intro stats class. But notice that our name, CourseKata, has nothing about intro stats or data science in it, because at heart, the co-founders are learning scientists. And what we care about is not just one course, but how a course improves over time using data from student motivation, engagement, and learning. “Kata” is an improvement routine in the Toyota Kata context; they used it to improve the manufacturing of cars, but we use the improvement routine to improve the teaching of courses.

Our kata is [based on] the insight that not everyone can do everything. Sometimes in higher ed, instructors are told: do everything, everywhere, all at once. Be everything to all your students, and do the research, do modernization, be relevant, do everything! That’s virtually impossible. So people end up either getting burnt out, or trying and failing, or trying and not bothering to measure if it’s true. But with CourseKata, we figured, “What if we brought the right people together to each do a small part that they’re really good at and very motivated to do? Then everyone wouldn’t have to do everything, and the courses could improve based on student data.” So we bring together researchers, instructors, designers, and developers to work together on improving courses at scale.

Lark Park: So what has the journey been like to not just launch, but mature? Maybe it’s never completely done, though, based on what you’re saying.

Ji Son: Every step of the way feels hard. When we were first pitching this idea to grants and different people, everyone thought it was a good idea. Sure, courses should improve at scale. 1.5 million college students take intro stats every year, and it’s a shame that we’re not learning and improving from that implementation. Every year they do it, and we have no data on what we should change for the next year. That’s a crazy waste! So everyone thought, “We should improve from students actually engaging in instruction. But ugh, that sounds like an almost impossible project—how do you start breaking down that giant task?” And we had a pitch!

So Chan Zuckerberg funded us to start this project. And then California Education Learning Lab funded us to see how it worked at three really different institutions—a UC, a Cal State, and a community college. If we make something that works super well at UCLA and then try to scale it somewhere, it’s going to be tied down to the various features of UCLA. But by starting with three different institutions, we built in the pain of having to scale across really different institutions from the beginning. We embraced variation from the beginning rather than seeing it as, “Oh that’s a headache that we’re going to deal with later.” That helped us get out to now 150 institutions, including community colleges, universities, even high schools.

Lark Park: I think you’ve gotten a lot of traction with high schools, which I didn’t know if you had expected that, or even had that in your sights at the very beginning.

Ji Son: At the beginning, we were just staunchly in intro stats. There’s enough introductory statistics sections in college to keep us busy probably forever. But some high school teachers were just poking around and saying, “Hey, I have a year-long stats course, can I use your materials?” And we were like, “Sure.” And that’s really how we got started in high school.

Lark Park: One of the things that I think has been very much part of your success is your professional development model. I’ve always thought it’s one of the best professional development models I’ve seen out there, and how it’s baked into how you think about community and improvement. So can you talk about the model, and has the approach and the model held as you’ve scaled to other institutions?

Ji Son: That’s one of the highest praises I’ve ever heard about our project, because instructor time is precious. Every hour of their time that we have, we want to make sure it’s high value. And what we’ve always thought is that instructors are not just out there implementing the stuff; it’s not like, hey, we have this thing, we’re just going to train you how to use it. From the beginning, the idea of the CourseKata model is that we all have a common vision: that all students will achieve flexible learning of really hard things. But our materials are not there yet, and that’s why we need instructors to help us get there.

So our professional development is called Study Group, because it’s partly about showing them the materials we have so far, but also letting them know: let’s figure out the things that don’t work and share ideas for how to improve it. It’s that interaction that is going to help us figure out how to make it work at scale, in diverse institutions, with different students.

We’ve seen a shift as we have scaled the professional development. We started with individual instructors randomly hearing about CourseKata and getting interested. And a shout out to so many of our instructors; they came into the CourseKata community by reading our research! That’s great for finding gung-ho early adopters.

But as we’ve grown, we’ve started working with whole departments, because sometimes the instructors who don’t go to professional development opportunities, like the adjunct instructor who gets handed a syllabus two weeks before class starts, those are the instructors who actually need professional development and support the most, but they’re not the people reading journal articles and deciding to adopt innovative curricula. So now we have a project where we’ve established micro-badges from the [CSU] Chancellor’s Office to train instructors whose department wants to use CourseKata and use research to improve the teaching of intro stats. We’re trying to support whole departments, and that’s a new developmental stage.

Lark Park: That definitely sounds like you’re evolving the model to accommodate scale. Does the ethos of “help us improve” still hold?

Ji Son: It’s always there, because it turns out the world continues to change; students are changing, the technology is changing, everything is changing all the time. So improvement is never over, and there are new elements of modernization. How do you prepare students in intro stats for the world of AI models? That’s a new emerging question that wasn’t on the radar ten years ago, but now is an urgent question. Because we have this continuous improvement methodology set up, we were already engaging with that in little ways—small steps toward it, but as a community; not just like, “Oh, we’re going to do a whole-hog redesign and roll out a whole new thing.”

Lark Park: Again, I think it’s a beautiful model, and one that’s probably too rare in higher education. But speaking of AI: what role has AI played in your plans for CourseKata? Do you see AI as having changed, or changing, what’s possible or necessary, in this learning context and improvement context?

Ji Son: I want to answer it at two levels. One is a question of how AI impacts the enterprise of education, of which we are a very small part. That’s one thread. Another thread is, I think AI is changing what’s necessary and the opportunities available in an intro stats class. I’ll start with that second one, since it’s a little more tightly bound to us.

AI models are fundamentally models trained on data, and because I’m a learning scientist, we’ve asked a bunch of people at scale—regular people at the beginning of an intro stats class, like a freshman college student—what do you think that means? And they say all kinds of crazy things, and a lot of it is not good from a perspective of wanting people to have AI literacy. People literally say “black box” so often that it’s become a code in our codebook. That is not the foundation upon which AI literacy is built. So how do we work on that problem at the scale of the entire nation? Doesn’t that sound like a real problem?

Lark Park: Not even a small nation, a pretty big nation.

Ji Son: This gives us a reason to look at intro stats differently. It’s not just this dumb GE [general education] course that students are forced to take, but all of a sudden it becomes national infrastructure that can be used for foundational AI literacy. Those pipes are laid—the fact that 1.5 million college students a year take it, and that this class is ostensibly about helping people understand data… Wow, I can’t believe we just had that lying around! Let’s use that creatively at scale.

So, with CourseKata, we had organized all of intro stats around the concept of modeling data, for learning science and data analysis reasons, because that’s what modern statistics is all about: DATA = MODEL + ERROR. But most students have no idea what that means, and a lot of our instructors are leaning into teaching intro stats as modeling. So CourseKata is now perfectly poised to ask students at the end of a stats class, “AI models are models trained on data—you just took an intro stats class, did any of what we taught you help you understand that?”

We’re just coding data on that, but at least from one class that we coded—I was not expecting students to write amazing things, we were just like, “Let’s just get started, let’s see what people say”—[around] 40% of the students in our very small sample said things that I would love to throw on a slide deck and show at any talk. They were saying things like: you want a model to capture patterns; you want to adjust the model so it’s able to predict things that are not in your training data, but predict data from the future, or data that’s outside your data set; and you want to adjust the parameter estimates. (Okay, very few people mention parameter estimates, but I know at least one person did, so I’m very proud of that.) Students bring in all these introductory concepts that are foundational to the world of AI, and that part makes me very excited. But can I now take a turn for a downer?

Lark Park: Yes.

Ji Son: I am also very concerned about AI, generative AI, for just the enterprise of education itself, and—

Lark Park: You’re not alone.

Ji Son: Yeah! Here’s the fundamental issue: we lived in a world where the markets needed deliverables, and in the past, people made deliverables. So the markets needed a system that developed people, and that’s the thing we called education. Developing people is slow and resource-intensive, but the markets were aligned, because they felt it was important to develop people. But now AI can make the deliverables, and the world still fundamentally cares about cheaper, better, faster. And AI can do it cheaper and faster, for sure. I have the sneaking suspicion that to be truly better in the future, you’re still going to need humans, but I don’t know if the markets can withstand the cost it’s going to need to bear to put up a system like education. It’s just too slow and too resource-intensive. Dealing with humans is not easy to do. I don’t know if that trade-off is going to be worth it.

Lark Park: Wow, that’s even grimmer than I was expecting.

Ji Son: Should I say, “You’re welcome,” Lark?

Lark Park: No!

Ji Son: Are you regretting asking me to do this?

Lark Park: No, I’m not regretting it, and that’s a very interesting systemic view. However, we imagine that the education of people is still going to yield things of value?

Ji Son: 100%! Here’s the fundamental thing that gets lost in the making of awesome, cheap, fast deliverables: you need somebody, somewhere, who understands the thing. Education, as a whole system, produced people who could make deliverables, but at the same time, it was producing people who understood stuff! And I would argue, just because you make deliverables with AI doesn’t mean you could offload all the understanding of stuff to AI. You’re going to need humans who can interact with legible systems and figure out when things go wrong; people who have good causal models of the world; and people who understand themselves and others as human beings. And if we don’t develop those kinds of people, you’re actually not going to get a lot of value from a system that quickly generates deliverables. Right now, we happen to be in a moment where people who are using AI systems already have developed a lot of understanding. But my question to the world is, “Can understanding emerge without doing? How much do you believe in that?”

Lark Park: That’s definitely the question, and I hear that being asked in very much that way, without an answer. But I think some people are trying to figure that out—genuinely, like, what’s the intellectual honesty of the answer to that question. But I worry that not enough people are trying to figure that out, and it needs to be figured out. And it’s gonna cost us if we don’t figure it out.

I’m reminded of the very first time we talked about ChatGPT, in December of 2022, when you said, “Have you heard about ChatGPT?” You even joked at that time, “Does this mean students will submit work done by AI to teachers who will have AI grade them? The humans being just incidental. Chuckle, chuckle.” And now that’s a key part of what higher ed is wringing their hands about. So, we’re three and a half years from that point. Have your thoughts about AI, and specifically large language models, changed in this window?

Ji Son: Generative AI was just a twinkle in people’s eyes back then, in 2022. We were so cute… It’s funny to have you recount to me the things I had said at the time, because you could also see my Ali Wong of data science angle working even then: it’s like, real serious end-of-the-world coming, let’s all laugh!

Lark Park: I think we did laugh at the time.

Ji Son: We did. Do I think it’s different? In a lot of ways, no. I actually even harken back to a time before LLMs came out—I’m terrible, let me just say that—I am a thorn in the side of the CSU at times, but that’s the beauty of being a faculty member with tenure, right? One of the things I told the Cal State system when they rolled out Graduation Initiative 2025 was: “You guys, if you focus on graduation, that is exactly the wrong key performance indicator. Because graduation and grades and passing are easy. You know what’s hard? Learning. Understanding. Can our students actually do stuff? That is much harder. And if you orient our whole system around that [graduation, grades, passing], we will rue the day.” I feel sad that I was right about that.

Lark Park: Elaborate [please!].

Ji Son: I think we have built our whole educational systems around passing and getting degrees and certifications, and not enough on the actual measurement of understanding and learning. That’s the thing I worry about. I think individual faculty care a lot about learning, but I think we all feel pressure to do things like increasing enrollment and increasing throughput. I think at best it’s independent of learning, but at worst, it’s at the cost of learning.

Lark Park: Do you then feel like AI has made us vulnerable, because we’ve chosen these metrics and oriented everyone toward a different type of motivation in a way that then really does hurt learning? I’m trying to get my head around the intersection of what you just said with our core vulnerabilities as educators in this time of AI.

Ji Son: That’s exactly it. I wrote a McSweeney’s article, a satirical piece on this new innovation in higher ed called Vibe Teaching™, and one of the jokes I made in there was, “All our dashboards show that we’re doing great!” That is the world I worry about: that the education system cannot even detect whether we are suffering in terms of student learning or not. And I have to shout out projects like PrairieLearn and UC Berkeley which has a testing center that’s fantastic. I applaud folks who, early on—before generative AI—just recognized, we have to get better at measuring learning. And we have to get better at that at scale, not just leaving it to individual instructors, but solving that problem in an institutional way.

Lark Park: Have we lost the plot in education? I do think we are struggling right now with what is effective in education, and what is an effective education; they’re not exactly the same question, but similar. I think your answer is, we’ve absolutely lost the plot.

Ji Son: That’s not to say there aren’t real bright spots of people trying not to lose the plot, and showing innovative, amazing ways of getting our system back on track. Definitely one necessary strategy is getting better at measuring the transformation of our students and being truly creative: to think about what they’ll need to thrive in every version of the future and how to figure out if we’re helping them get there. But I don’t think measurement is enough.

Can I give you an example? There was an experiment done with kids in Texas, NY, Chicago, by a Harvard group, where they paid kids for better grades in one group versus another group who got paid a dollar for every book they read. So one is rewarding the outcome versus the process. Can you guess which was better?

Reading the books seemed to be causal in helping improve the grades because it’s not the motivation to improve the scores that is the problem. A lot of the time, people don’t know what to do, even if they want that outcome. And I think the same with our students. We definitely need to get our measurements of the right things in line, but that’s not going to be enough. You can’t just say, “Hey students, you’re failing at this. Figure it out. Good luck!” We have to design and improve learning opportunities where students could engage in the enterprise of developing and transforming.

Lark Park: So I want to bring up math in this context, because this is one area where what’s being communicated broadly is: they’re failing in math. And CourseKata also got caught up in the recent math skirmishes in relation to Area C. [Area C refers to mathematics requirements for admissions eligibility to UC and CSU. The controversy was centered around whether there was a “loophole” related to which courses could count in lieu of or “validate” the Algebra 2 requirement in Area C.]

Ji Son: When validation was first proposed, the idea was: let’s say you take Algebra 2, but then you go on to take pre-calc and calculus, and you pass those courses. Pre-calc is like a review of Algebra 2, so that makes sense that if you pass pre-calc, you don’t need to go back and take Algebra 2.

Statistics doesn’t [cover a lot of Algebra 2], not because it’s not rigorous, but because it’s just a different domain; it’s really focused on statistics concepts. It’s a different fork in the road. Courses like financial math, discrete mathematics, AP statistics, and any other math course that diverges and goes on a different road than Algebra 2—none of those courses should validate. That, to me, seems like a clean policy solution, [but] I think what they ended up doing was just picking three curricula [of which CourseKata was one] and singling these “data science” courses out as uniquely terrible at validating Algebra 2. Yes, we are terrible at validating Algebra 2. What I argue with is the “uniquely” part.

Lark Park: That was a confusing period. What it didn’t resolve were the concerns that students are not doing well in math, period.

Ji Son: We still have not solved the problem we started off with, which is that students are not motivated to learn math, and they really struggle in learning.

When we say students don’t pass Algebra 2, that doesn’t tell me anything about what concepts they are struggling with. If we could get back to the questions that are closer for teachers and closer to the students, I think that would help our whole system. But because the measurements we have are often at the level of things like “passing,” it’s hard to make progress on the problem of learning.

Instead of orienting around what classes are called, wouldn’t it be great to have a system that can incrementally adjust to improve this outcome: whether students can apply math concepts to a variety of situations. The education system needs to learn, not just the kids. That’s the vision I would like us to get behind.

Lark Park: I did want to ask you [about] your other projects—like the romance novel [you wrote] and [your personal interest in] housing policy—whether there’s bleed-over [to teaching statistics]?

Ji Son: There’s always bleed-over. I wrote a romance novel called The Long Con—which really started off as a Google Doc that [I] sent to my friends but is now available on Amazon—and that led me to pull some data about the financial impact of romance scams from the FBI. These [confidence] scams are a global industry that’s bigger than the illegal drug industry. So now we have a student project where intro stats students can examine confidence scams that literally involve the whole world.

Lark Park: You’re turning this into a pedagogical project?

Ji Son: That’s right. It’s all in service of learning truly important and interesting things! And we also have a housing policy lesson that helps students understand scatter plots, “correlation is not causation,” and how homelessness and housing policy might be interlinked.

Lark Park: That’s amazing.

Ji Son: And at the end of it is a call to students to get involved civically, to go out there and help collect data on housing issues because that literally requires people walking and driving around their congressional districts every January. Cal State LA students go every year and participate in our local homelessness count, and also do things like call their city, state, congressional and senate representatives. So many of our students have never called a representative in their lives. If your statistics class gets you to call your representative, I’d say that’s college producing externally valid learning outcomes.

We Must Address the Growing Rage Against the A.I. Machine

July 11, 2026, NY Times, Eric Schmidt and Selina Xu

Yet how technology spreads is never inevitable. If A.I. is viewed as benefiting the few at the expense of the majority, then the public will rage against the machine. And A.I. won’t be able to make our lives better in the long run if it cannot survive in the short term. The real challenge, then, isn’t whether the United States or China will build an overwhelming, insurmountable advantage over the other. It’s whether either can figure out how to realize the benefits of A.I. without ripping apart its social fabric. Neither has found the answer yet.

Economic Futures in the Age of AI

May 26, 2026, Open AI Foundation Announcement, Divya Siddarth and Wojciech Zaremba

Many current approaches to studying AI’s economic impacts focus on which tasks could be automated. This is useful, but incomplete. The economic effects of AI will depend on how tasks are bundled into jobs, whether automation displaces human labor or creates new labor-complementary roles, how task distributions shift as model capabilities improve, and how firms and states reorganize around those changes. Understanding these shifts requires better labor market public infrastructure worldwide: BLS-like capacity to measure employment, wages, transitions, and firm behavior, alongside modernized O*NET-like systems for mapping work. These systems should be globally relevant and linked, where appropriate, to demographic, geographic, career-stage, and job-level information.

Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build

May 20, 2026. arXiv preprint, Sina Rismanchian et al.

Perhaps there are two practical implications of our work, one at the policy level and the other at the pedagogical level. At the policy level, our findings suggest that proctoring may help preserve authentic behavioral signals: the post-ChatGPT divergence in response times between AI-susceptible and AI-resistant topics disappears entirely under proctoring, while the same estimator applied to non-proctored retention items yields an opposite-signed effect. However, mandating universal proctoring may be impractical and inequitable, as the costs and logistical burdens of proctoring are likely to fall disproportionately on under-resourced institutions. A more scalable response lies in the deliberate design of learning tasks. Our results suggest that problem formats that are inherently resistant to AI assistance can preserve authentic engagement without requiring supervision. Increasing the proportion of AI-resistant problem types in pedagogical design does not depend on proctoring infrastructure and does not impose additional surveillance burdens on students. That said, the growing prevalence of multimodal foundation models and agentic AI systems, such as models capable of controlling browsers and operating systems, means that the boundary between AI-susceptible and AI-resistant problem formats is a moving target. This boundary will need to be re-evaluated as multimodal and agentic AI capabilities continue to advance.

Generative AI tools now sit adjacent to nearly every educational task a student undertakes. Our findings are strikingly alarming: students are spending substantially less time on AI-susceptible problems, and this shift is associated with a substantial decline in retention of the underlying concepts. If we aim to foster human learning in the generative AI era, the way we inform them about AI use, design learning tasks, assess student progress, design AI policy, and infer mastery will need to be rethought…

Learning Lab Director Lark Park interviewed Dr. Kirst in a wide ranging conversation – from how education policymaking has changed over the decades, to whether UC should bring back standardized testing, to why the professional development of teachers, the structure of governance, and alignment of interests all matter for educational success.

The conversation was edited by humans with help from Claude.ai.

Lark Park: Mike, you’re kind of a legend in many education circles. You’re one of those rare people who cross over between K-12 — really TK-12 now — and higher education, and you cross over both as an academic and as somebody who has deep experience in government. You had a lot of responsibility, at least on the state side, and I think the federal level as well. You’re still the longest-serving president of the California [State] Board [of Education]?

Michael Kirst: Nobody will top it, because nobody’s had 16 years in office with a governor. Even in the past, it was unheard of to serve 16 years.

Lark Park: You’re going to hold that record for quite some time, if not in perpetuity. But there’s so much you’ve done in so many different circles…. What do you think has been the most impactful part of your career?

Michael Kirst: I think the starting point of my career was impactful. I was there at the beginning of the Johnson administration in 1964, and I was 24 years old. By the way, how I got into this field: the dean of the Public Policy School at Harvard — when I got my PhD in political economy — advised me to go to the Office of Management and Budget to start my government career. I never intended to be a professor. I wanted to save the world in Washington. He suggested that, and the Office of Management and Budget had openings in veterans, water pollution, and K-12 education. They said, you can have any of those, and I said, well, out of that lot, I’ve been to school, so I’ll take K-12 education.

I had gone to graduate school right next to the Harvard Education School and never walked in the door. I wish I had. If I’d done it over, I would have walked in the door. That’s how I got going. There was such a small governmental decision-making operation [at the time], and so I really had a big impact on Title I, the largest program. I remember talking to Lyndon Johnson when I was 25 years old in his office about how much we should spend, and what the formula was.

Lark Park: Wow.

Michael Kirst: It was unbelievable. I then came back as Director of Program Planning in the Office of Education. I wanted to get closer to the programs, and so I wrote some of the initial regulations for the first federal programs. We were the only ones doing much in the ‘60s. The states were quiescent and not really moving much at all, and of course, they were segregated at that time.

Then the Johnson administration ran out of gas over the Vietnam War, so I shifted to the Senate. I became the staff director, and they called me Counsel of the United States Senate Subcommittee on Employment, Manpower, and Poverty. So I got this overview of employment, post-secondary, and job training. I was crucial in putting through a bill called the Manpower Development and Training Act. So I had not just K-12, but all forms of post-secondary — not just college, but job training, now called workforce. That’s how I got to span the levels.

I would say that has been an important part of my career. The work that Andrea Venezia and I did around 2000 — we published a paper called “Betraying the College Dream: How Disconnected K-12 and Postsecondary Education Systems Undermine Student Aspirations” — that began the conversation about students not completing college and not making the right choices, and so on.

So, I think I had a big impact there. Then with Governor Brown, the local control finance formula moved us from 45th in spending to — if you don’t adjust for cost of living — 13th in the country. If you adjust for cost of living, we’re above the national average, and we are now rated the second most equitable system of school finance of any American state. We were down in the 30s and low 40s when we came in, so that, I think, speaks for itself in terms of moving in the rankings.

Lark Park: Where you started your career explains so much about your ongoing interests, your views — just the perspective you have is much broader than a lot of people who we work with. I wanted to get your perspective on some of the biggest challenges you’ve faced in policymaking. How did you overcome them?

Michael Kirst: Only in the most recent years would I answer this differently than before. In my earlier career which goes on for many decades — I was basically focused on policymaking, and that often featured finance, government, governance, and categorical programs for specific purposes.

A lot of the work was on how to construct winning coalitions. I taught the politics of education at Stanford every year I was there, from 1969 through 2007. I also taught students how to think and act like a politician. At one point in my career, when I did that Senate job, I worked for a Pennsylvania senator and thought about going back and running for Congress. But he was defeated, then Nixon was elected, and I left to teach at Stanford.

In the early period, the challenges were understanding the politics and how to put together a winning coalition — and, as an academic, how do you speak truth to power?

I co-created, with Jim Guthrie, an organization called Policy Analysis for California Education, which now leverages five leading universities in California to do education policy and apply it to decisions coming up in the short range in Sacramento. The challenge was tailoring academic work, reframing it, and making it so that policymakers could grasp it. That was, I think, the big challenge.

The biggest challenge now — and one that I confronted with Governor Brown — is how do you connect the capital to change in the classroom? That is the hardest problem in American K-12 education. Our mathematics scores, for example, are very poor across the United States. How would you use policy to change what the average 4th grade teacher teaches in mathematics, and provide him or her with the capacity to do that? As policymakers, we have been able to build things in addition to the classroom — add-ons like advanced placement or career and technical education — but we haven’t been able to change how the basic subjects, like literacy and math, have been taught. That’s a challenge I’m still struggling with.

Lark Park: You and a lot of other folks. I want to talk about math more, but when you said “capital,” I found myself thinking: did you mean capital in terms of money, or what goes on at the State Capitol?

Michael Kirst: State Capitol and federal capital to the classroom.

That was what we faced in Title I way back in 1965. Could we really improve classroom instruction? That’s the real dilemma now. Policymaking has turned a lot more toward improving instruction, improving learning, and improving the capacity of local classroom educators and central office people to align their policies and get more change.

The weakest link in our whole system, to me, is that we can influence what teachers teach, but not how they teach it. We’re trying to put in a bold, new curriculum in mathematics in California featuring conceptual understanding. How do you teach conceptual understanding when you’ve been teaching formulas — solve for X, memorize the various formulas, plug in numbers — often without understanding the conceptual underpinning?

Lark Park: Do you think that’s still a struggle, but at least we’re looking at the right end of the telescope?

Michael Kirst: Yes, I do. You’ve seen the science of reading, and 40 states have adopted it. That is generally the right direction.

I’ve studied two places in depth. One is Ontario, Canada, which did this — changed out, I think, around 80% of their teachers between roughly 2005 and 2015. And then, of course, Mississippi. There is proof that it can happen. But how does this happen in a behemoth like California, which has 375,000 teachers and other instructional aides, at that scale, across such a huge geography? Mississippi is a much easier state to work with.

Lark Park: I want to stay on policymaking for another minute, because I understand that the approach has changed, but can you elaborate on what is different about the education policy environment today compared to, say, 30+ years ago?

Michael Kirst: There has just been a recent study of that. I completed my last textbook on the politics of education — called “The Political Dynamics of American Education“ — in 2009, and wow, has it changed. The biggest change is that there are many more interest groups entering the public education system, and they have begun to weaken the influence of teacher unions.

There are all kinds of parent groups, ethnic groups, and cultural groups. There are a lot more players, and it’s a lot more complex and more pluralistic in terms of politics. It was a lot easier to describe even as recently as 2009, when there were fewer interest groups and less ideology. Numerous low-income and ethnic groups have entered the field as well. So it’s more difficult to put together coalitions, because there are so many more interest groups to deal with.

Lark Park: Why do you think that’s occurred, and do you see it as a positive change or a negative change?

Michael Kirst: I think it’s a positive change. Democracy is better and more flourishing when you have a lot of different interests bringing their viewpoint and representing a much broader array of citizens. Why it’s happened is that people have become more dissatisfied with the results — either the cultural results that the governor of Florida features, or the results for low-income and minority students that organizations like Education Trust feature.

They’ve also confronted the employee groups, who have their own needs — cost of living increases, and salaries not keeping up with other master’s degree-type graduates. You’ve got a much richer array, but I think we’re looking at the problems in a deeper way, and it’s being produced by a bottom-up democracy.

Lark Park: I can definitely understand why that could be good, but you’re saying that part of this is a reaction — to being dissatisfied with what our educational system is producing, what the outcomes are. You had mentioned math test scores as well, and there’s a lot of dissatisfaction with math scores in particular, literacy as well. (COVID hit everybody hard.) But you recently co-authored a paper challenging this narrative of educational decline. Tell me, what motivated the paper? What led you into this inquiry?

Michael Kirst: When we talk about education decline, it has been mostly about elementary and secondary education in terms of student outcomes — the national assessment and test scores driving the narrative, along with attendance declining. K-12 was hit hard by the pandemic, and that explains a lot of it, but there was a lot of dissatisfaction before the pandemic.

All the work done in looking at the seam between K-12 and higher education, and all the problems students were confronting when they weren’t completing college — combined with a changing workforce and a changing economy where people needed to know more — led me to see that the national discussion on K-12 is about decline. But what about post-secondary education?

Very few people have no education beyond age 17, and the [last] test scores are all [measured at] ages 15 and 17. If you’re making up some of that ground after graduation — what does it look like when you’re 25? You’re still regarded as youth, and I felt there was progress there.

We hadn’t included the frame of people in education from preschool through 25. That was the motivation. People don’t think across the system — they write about K-12 or higher education, but they don’t put them together. What we found in the paper was a more positive picture in post-secondary education that you could make a case for offsetting the decline in K-12. To what extent is not clear, but I don’t think you can say that post-secondary education has been declining in terms of years of schooling. My dependent variable was the years of schooling, and those are going up.

Lark Park: I do think there’s something to be said for how we might look at thriving across a lifespan. This is something your colleague Mitchell Stevens with Learning Society is thinking about.

Michael Kirst: What I was concerned about in the paper was that it may be better than you think, and we need to really understand what’s going on for youth — rather than just measuring with a test score at a fairly early age.

The tests point to a problem, but what happens next is important. I tried to provide a framework of 11 sectors of post-secondary education — everything from what prisoners learn in prisons, to what military people learn, to college degrees, and all the certificates, licenses, credentials, and badges. We’ve had an explosion in those.

Lark Park: One of the points in the paper is how these areas get neglected in terms of measurement and evaluation, and they [the neglected areas] could be quite productive. It makes me think about how, working in higher education for a while now, it’s college, college, college — and college only lasts for so long. There’s a lot of life to live after college, and a lot of education in that — experience in different jobs, etc. It [college] is an important and formative time, but there’s so much more out there that we’re not looking at. As a system and as a state, that’s something we really ought to do more of.

But let me focus a bit on test scores in K-12, specifically math. You made a distinction between procedural math — what we’re teaching — and conceptual math, which we should be emphasizing more, not to the exclusion of procedural, but enriching it. Are we measuring the wrong thing when we test students in K-12?

Michael Kirst: Yes, I think it’s partly true that we are. The math curriculum sequencing — I graduated from high school in 1957, and it hasn’t changed. It is still driven by calculus, and therefore everything is linked to calculus. Not many students really pursue calculus; maybe more should, but that’s too narrow a frame. What about data science in high school? What about statistics? We all use data science and statistics a great deal.

The sequencing is all aimed at calculus as the nirvana. That’s why you have to start in 8th grade — and the big argument in Palo Alto is whether to offer advanced calculus. In many ways, we haven’t understood the shifts in what kinds of quantitative analysis students need, but we can’t really break the hammerlock of this calculus-driven system.

Lark Park: One of the things I wanted to ask you about is the Common Core State Standards for Math. You were instrumental in the state’s adoption of Common Core. I remember a conversation we had — you saying that adoption was one thing, but the professional development and implementation were something else. Those got shorter shrift. Now, being [more than] a decade out from the adoption, what is your perspective on that?

Michael Kirst: It’s still a problem. Let’s start at the beginning. Almost all states stop their teacher responsibility once the teacher is credentialed. We have the California Commission on Teacher Credentialing, and they stop after the teacher is credentialed. They have no more responsibility. It’s as if you’re inoculated for your career when you get the credential when you’re 22 years old or a little bit older. There’s no vision that it’s the state’s responsibility to ensure the capacity of teachers to adapt to changes over time — in curriculum and what people need to know. And then our professional development that is carried on during teachers’ careers is very short, fragmented, and not sustained enough to really enable [for example] a fourth-grade teacher to teach a new mathematics curriculum. We’re going to totally fail with the new math curriculum here unless we have the professional development.

Under my watch, we created islands of excellence and huge deserts where no teacher capacity had been built to implement the Common Core. We never had a strategy to enable our existing teachers to teach it. Teacher education institutions are slow to change their preparation.

Lark Park: With regard to math education, we seem to be in this real period of angst, [with] different reports coming out. It’s gotten to the point where [many] faculty at the UC system have basically said, bring back the SAT, we need tests that really tell us when students haven’t learned math. I’m wondering if you have an opinion about this approach, which curiously does not focus on how to bolster math education, but rather just on finding out who knows it and who doesn’t. (You know that I’m on the UC board, so this will matter at some point.)

Michael Kirst: I’ve never been a big fan of how the SAT was carried out by the College Board and their contractor, Educational Testing Services. So let me frame it this way: do we need a good test that tests the right kind of mathematics? I’m open-minded to that. I haven’t made up my mind, but I would not go back to the old SAT and just use that again. People should have a higher vision — we need to move beyond the SAT to a better measurement. Our California test, the Smarter Balanced Test, has problem solving in it and is much more conceptual than the SAT. I’m against bad tests, and I’ve never thought the SAT was a good test.

If the leaders of higher education could say, let’s open it up to what a good test would be — let’s acknowledge that not everybody should be tested on whether they’re on a pathway to calculus. So, I would want a test that has much more about quantitative reasoning, statistics, and various kinds of analysis — new forms of computer science — rather than the same old, same old.

Lark Park: It’s certainly a time to rethink what students should be learning, and how they demonstrate it — related to what pathways. There’s still tension about what all students need to know versus when they should start branching out into different meta-majors. I don’t know that we have any sort of consensus about when that should take place, so this is a great time to re-examine what students need to know in order to succeed in college or career. What disturbs me is when very smart people ask: with the advances of AI, do we even need to know math? I think on one level they’re joking, but I’m not actually sure. What sort of conversation is there in the circles you frequent, in terms of AI’s impact on education and what we need to teach?

Michael Kirst: First of all, there’s a profound belief among a strong faction of people that it’ll be a while before AI really affects K-12 schools. I grew up when radio was going to change it, then TV was going to change it, then desktop computers were going to change it. We had the Apple Classroom of Tomorrow with big, bulky desktops. Then I was told that when we get handheld phones and small devices, that’ll really change it.

That hasn’t happened. Now we have Chromebooks, and there’s an all-out revolt coming from the bottom up — interest groups leading the push to get Chromebooks out of classrooms. Technology confronts the classroom, and the classroom wins.

The next thought I have is this: I was around when they announced the Apple II in 1982. The Apple I was built for industry, and we conceived that there was a lot going on, but we had no idea what the current array of Apple products would look like. So if you’re asking me in 1984 what the impact of the Apple II would be — I can’t imagine it. The answer is I don’t know. But one of the last places AI will probably come in is the K-12 classroom.

It’s different in college. For example, I just had lunch with the head of Foothill De Anza Community College here in California. Sixty percent of their students are fully online, so it [AI] has [the capacity to] penetrate higher education in a much more robust and deep way. When you ask me about the impact in education, I’m going to split K-12 — and now preschool through K-12 — off from higher education. I think it’ll be more profound in higher education, though I’m not sure how, because I don’t think we can imagine what we’re confronting. K-12 schools will be at the tail end, in my judgment.

Lark Park: Maybe that’s a good thing. There’s definitely a growing contingent of parents and likely teachers who think it’s good that K-12 will be a holdout against AI. But there’s also a sense of encroachment — companies selling their wares and promising the next greatest thing. There’s a lot of healthy skepticism there.

Talk to me about the latest move to take the Department of Education and put it under the State Board [which narrowly got approved in this year’s budget negotiations]. I know this must be of great interest to you.

Michael Kirst: I’m an ardent supporter of it. The system now is so fragmented and misaligned — the state power is so fragmented across the Department of Ed [CDE], the State Board, and the California teacher [credentialing] commission, which is separate as well. Putting the State Board and CDE together is essential.

The board’s mandate was to make policy. We had pretty good policy — I like the Common Core, and we still use it. But the implementation was over in the department. My role ended at policy, and the department was responsible for implementation. So neither I, nor the board, nor the governor had much ability to influence the department. It’s an independently elected official, not part of the governor’s cabinet.

You can’t separate implementation from policy. You can’t influence the classroom from the capital with a fractured, misaligned system.

The California Department of Education [also] has trouble recruiting people, and it tends to elect legislators, board members, or people who haven’t really run schools — so they don’t really understand it. Under the new [structure], there would be a commissioner of education who would hopefully have seen and done all this, could be from out of state, and would be much more highly paid. The department could then begin to attract newer and better people with higher pay.

Lark Park: So what it might fix is the disconnect between policy and implementation, because they are currently separated. But even with such a move, what are the things this won’t fix that we still have to figure out?

Michael Kirst: Organizational change merely creates a platform and a structure that may enable 10 more different things to really improve, for example, the math curriculum taught every day in classrooms. It’s no silver bullet, but it helps. It helps get you started, helps with leadership, and hopefully helps with a strategic plan to get from here to there. Going back to my fourth-grade math teacher, often not well prepared to teach math and better prepared to teach literacy…that’s a long conversation. I think there are 10 buckets, and governance change is one bucket.

Lark Park: Maybe it’s that first step in creating the kind of alignment necessary for further change.

Michael Kirst: That’s right, and we could get better if we could align our curriculum in different ways with post-secondary education.

Lark Park: That’s a project the California Education Learning Lab is working on, actually — that alignment, not just of curriculum. I think you noted [in your paper] that Jobs for the Future has a proposal to merge grades 11 through 14, and to think about that time period — ages 16 to 20 — more cohesively, rather than having this big divide between ages 17 and 19 [between secondary and post-secondary].

Michael Kirst: That’s where the thinking has to go. You’ve got to have a lot of parts changing, and combining the state board and the department doesn’t do anything to bridge that gap. My biggest frustration in my whole career is how separated K-12 is from higher education, and how we are organized in very different professional groups.

When I would speak about the problem of transition from high school to college, I rarely got an audience that had both higher education and K-12 people. They don’t meet together. I’ve called it an unnatural act performed by unconsenting adults, trying to get them to come together. I could never even get an audience of both in the same room. All the subject matter people are pretty much isolated from each other — K-12 versus post-secondary. We made progress from around 2000 to recently, but it has stalled out. We’ll see.

Lark Park: It’s like the pace of continents moving. Is that a project you’ll likely be working on? Any projects you want to let us know about?

Michael Kirst: Right now, I’ve got a very robust and positive response from this paper, and I’m trying to figure out what to do next. I’m at the beginning of a new endeavor to bring the levels together, rather than focusing on college preparation, which was the old focus.

The paper covers two areas that merge very well: advanced placement and dual enrollment. We should be proud of the increase there. We have produced a lot of college education in high schools, and it’s growing fast. If you actually look at our test scores — and nobody mentions this much — the higher end of our tests are going up. I keep saying, they’re all our children — we shouldn’t just look at the average test score. Yes, the bottom is going down, but the top is going up. So, we’re doing some things quite well, and we’re merging high school and college — dual enrollment is a really important outcome. That’s one of the positives for the two levels getting together.

Lark Park: Last question: You talked about the politics of education and the secret of constructing winning coalitions. For anybody working in this space, what’s the answer? How do you construct winning coalitions?

Michael Kirst: You have to envision the full field and try to figure it out. You have to understand the array of interest groups that will weigh in. When I try to get legislators to introduce bills, they say: “Who wants this? Is it just academics? I’ve got to have more than that.” You need to put together a wide-ranging group of people.

Guest:

• Dr. Julie Carpenter: learning scientist, author of The Naked Android: Synthetic Socialness and the Human Gaze and Culture and Human-Robot Interaction in Militarized Spaces: A War Story; Senior Fellow (External) with the Ethics + Emerging Sciences Group at California Polytechnic State University

Also available on: Apple / Spotify

CHAPTER 1 [00:00-3:10]

Dr. Julie Carpenter [Cold Open]: “The systems are going to be continually redesigned to be more immersive and engage and retain people. AI literacy is great, but it’s really hard for a course to keep up.”

Taiyo Inoue: Welcome back to My Robot Teacher. I’m Taiyo Inoue.

Sarah Senk: And I’m Sarah Senk. And today we’re talking to Dr. Julie Carpenter, author of The Naked Android: Synthetic Socialness and the Human Gaze.

Taiyo: Julie’s a researcher whose work examines how people make sense of artificial intelligence, robots, and other emerging technologies, and how those encounters shape identity, social relations, and ideas of agency. Her research focuses on the cultural meanings people bring to intelligent systems; how those meanings inform public expectations and design; and how people respond to, behave with, and judge these technologies in practice.

Sarah: One of the reasons Julie’s work is so useful for thinking about AI in education is that she does not treat human responses to technology as either purely technical or purely psychological; she is interested in the whole situation: the design of the system, the cultural stories surrounding it, the setting where people encounter it, and the meanings users bring with them.

Taiyo: Right, and so this is where the term “sociotechnical” comes in. Sarah, you kept using that word when we were preparing for this episode. What is a sociotechnical approach?

Sarah: I know, sorry for my jargon! I asked Julie how she explains the term ‘sociotechnical’ to people outside of her field and she said that the fundamental idea is that “you can’t pull the tech apart from the context it’s sitting in. You can’t point at some line of code and say “that’s why people get attached,” and you can’t point at the user and say “that’s just them projecting.” You have to look at how the whole system and the specific situation come together, and that means thinking about the collision of design choices, the institutional settings, the cultural stories people already have about AI, the economic incentives behind the platform, human psychological needs, and so on. So basically, a sociotechnical approach asks how the technology and the users of that technology, and the broader systemic conditions at play are all shaping each other simultaneously.

Taiyo: So today’s episode asks: When institutions adopt AI at scale, are they implicitly telling students it’s authoritative and trustworthy, and if so, how do educators reclaim their own authority in the classroom? If these systems are engineered to engage, validate, and emotionally resonate with us, is awareness and education enough to resist all of that, or does the power of design outpace what education can teach? And finally, what kinds of learning depend on being seen, challenged, and corrected by another person, as in another human person - and are those irreplaceable - in particular by machines? Now onto the episode.

CHAPTER 2 [3:10-10:26]

Sarah: Julie, hi, thank you for joining us. It was great meeting you at Cal Poly’s AI symposium in San Luis Obispo last month, where you gave one of the keynotes. And I realized as you were speaking that I had just read your recently published article in the journal AI & Society because ChatGPT Deep Research recommended it to me. And so, I have to thank you for being so patient when I bumped into you in the hotel lobby later that night and started excitedly rattling off all my thoughts about it.

Julie: What you didn’t know about that scenario, I had locked myself out of my room when I went to get ice, and when I bumped into you as I was getting into the elevator and I was thinking, “I just want to get my ice water and go to bed.” And you said, as the doors were closing, you said, “Hey, I saw you give a talk today. Great job.” And I said, “Hey, thanks.” And you said, “And I read your paper” and I stepped out of that elevator so quick because if you’re not in academia, let me explain that, like, you feel like you write, you work so hard on these research papers, and you work so hard to get them through peer review, and then they’re published, and then they’re rarely read, you know? So you said you read… And I never asked you how you felt about it ‘cause I was just happy that someone had read it.

Sarah: I did love it. And I also thought it was really interesting how I came across it:I was updating this bibliography articles about AI that I liked and wanted to save. And I, you know, I put it together maybe six months earlier. And I thought, oh, let me see what new stuff ChatGPT’s Deep Research can surface based on what I collected already. And I don’t even remember exactly how I prompted it. But it was something like, “I’m really interested in social constructivism and sociotechnical approaches to AI literacy, and so on - and so I gave it all these details about me, like how I work on grief, loss, and technological mediation. And your article was at the top of the excel sheet. And I don’t know, I think it’s also funny that from reading that paper and then meeting you and learning more about your work, I was thinking, oh my God, we have so many of the same interests, and I feel like this might be my first AI mediated or AI initiated human relationship, like finding somebody who shares a lot of my theoretical interests and academic predilections, right?

Julie: Yeah. We made friends via ChatGPT. So thank goodness I locked myself out of my room and my ego wouldn’t let me get off the elevator.

Sarah: [laughs] So, before we start the official interview - Julie, I really want you to contextualize a comment you made just before we hit record about how ironic it was that you couldn’t find the right microphone connector even though you, quote, “had more cords than a Radio Shack.”

Julie: I’m gonna make a lot of comments that will date me and the eras that my technology love formed in - with comments like that. My brother actually managed a Radio Shack, and for those that are unfamiliar, Radio Shack was, uh, like a mall store that specialized in electronics, everything from batteries to obscure cords, to things so you could DIY home electronics, build your own early computers, and purchase things, and kits, and toys, and it was like a very DIY-heavy Sharper Image kind of thing. Yeah. And you could just go in there and wander for hours. At least I could as a kid. But yeah, my offhand comment was a joke about the, the back part of Radio Shack that would just be full of bins of cords and- ... and, and things, right? Circuits and, you know.

Sarah: I know everybody is nostalgic and thinks that, like, the time they were growing up was the best time. But I, I feel like the time when, like, being, like, an adolescent in the time when RadioShack was popular, it doesn’t really get better than that.

Julie: I think it’s sort of an interesting downfall into like, home electronics, really. It was such a great entry point for kids…

Taiyo: Yeah

Julie: …and adults to nerd out and find bits, before the internet where you could just, you know, find a chat board or go to Facebook or your social media of choice and say, “Where can I find this?” It’s more than nostalgia. I think RadioShack was actually really formative for a lot of people. And for me, my father, and I’d mentioned this to both of you before, he was very into technology and electronics, and he brought home, you know, DIY circuit boards and things for my brother and I to build, so I knew at a low level what some of those things were. And to have the freedom to go in and, and, and buy little pieces of things so I could, you know, make my own circuits and play with the circuit board in different ways. I don’t know how kids replace that these days. I mean, they’ve got different toys, and there are still older toys that you can tinker with. I’m sure, like, there’s all kinds of literally tinkering toys out there.

Sarah: Yeah.

Julie: But the RadioShack was a unique experience, so may its memory be a blessing. Long live RadioShack.

Taiyo: [laughter] What an interesting way to open a podcast.

Sarah: Yeah, well I’m curious about this from a learning sciences perspective. though. As a kid, I was kind of a bystander watching friends play with circuit boards and early computers, but I also felt like they weren’t really for me. So I’m curious what makes a technology feel inviting instead of intimidating, and how do those early encounters with technology shape what people think they’re capable of learning - or even how they think and behave around technology later on?

Julie: That’s a big question coming off of RadioShack! You know, I do believe that RadioShack had a lot of great things that supported and scaffolded, which is a great Vygotskian word as a learning scientist I have to throw in. It scaffolded learning for kids of all, and adults of all ages. It was great because you had the tangible, available things you could touch, pick up, examine, figure out how they worked together or didn’t. You had the mentors available if you wanted to talk to them, the salespeople. And people, I think, were less afraid of direct one-to-one interactions than people are now in stores especially, and people don’t expect that kind of knowledge and expertise from, like, a mall salesperson necessarily today…

Sarah: Right

Julie: …the way you did at RadioShack. That’s what they were known for. So that would be in, like, learning scientist parlance, you know, like a guided practice. You know, you can ask questions, you can say, “This is what I was trying to do at home,” and the mentor can, you know, say, “Well, I would try it this way,” or, “I’ve done it this way and it works well.” And then you had the opportunity for independent practice, where you could take home these little, again, inexpensive… so it didn’t take a huge amount of socioeconomic privilege to get these little pieces together and, you know, do even the simplest DIY things, you know, make a potato radio or something, you know. Very low barrier to entry to figuring out these sorts of things. So, yeah.

CHAPTER 3 [10:26-19:10]

Taiyo: So I wonder if you can then comment on how those early formative RadioShack experiences contrast with the current day, um, and the way that tech, in particular, artificial intelligence, which is, of course, the big new tech on the scene, um, what that experience might be like for young people these days and thinking about it particularly from a learning science point of view.

Julie: Yeah, there’s different ways people can adopt conversational and chatbot AIs into learning, right? Some of the policies are technically all over the place school to school. Some teachers might require it for their course to a certain degree to teach kids critical thinking. Some teachers require it without critical thinking. Uh, some schools require it for lessons or that it’s a required course, so people don’t have a choice to opt in or opt out. And so the whole idea of these institutions adopting AI very rapidly is also an implicit message to the students and the learners that this is something that has been vetted by these institutions and adopted and accepted, and therefore it’s a knowledge system that should be relied on with a certain amount of authority because that’s what their teachers are telling them, or that’s what is implied by the school adopting AI or requiring courses in it that you must take.

Taiyo: And of course, this isn’t an abstract thing in the Cal State University system, which provides ChatGPT Edu accounts to all students, faculty, and staff - and in fact, recently renewed the contract with Open AI for another 3 years. So I guess the question is, once something like ChatGPT gets built into the learning environment, what foundational things should educators be thinking about?

Julie: I believe learning is fundamentally relational. Uh, that’s a real problem with AI. You know, you could say that it could go back and forth in things like, I don’t know, maybe Duolingo might be an example of going back and forth and modeling and guided practice. But the thing is that when you design AI-guided practice, it’s optimized for knowledge transfer, and it bypasses that whole idea of the relational dimension, that can undermine long-term learning outcomes - the whole idea of the person modeling what needs to be learned in a challenging way to the student or the person being educated, but doesn’t challenge them enough that it turns them off or creates anxiety and distress. And that’s something that a good educator should do in person where AI is not so good at. And I would have one other thing to add that is hard to disentangle from AI, and that’s the idea of surveillance is really hard to disentangle from it. Um, ‘cause any time something is heavily personalized, like chatbots or something that’s going to remember your history, people need to adapt to it. They learn to become legible for the machine and the way it reads people’s reading output and what it accepts and doesn’t accept. And they know to a certain extent that everything that they share on whatever the space is, Canvas or whatever tools are being used by an institution, um, could potentially be retrieved.

Sarah: You know, the thing I find most interesting is the stuff we don’t even realize we’re disclosing. I mean, Canvas gives instructors so much information about students - like how much time they spend with our course materials open, when they submit stuff, how often they revisit materials, you know, whether they seem to do everything at the last minute. And maybe one data point doesn’t tell you much but across courses and semesters, I imagine there are a lot of inferences that could be made about someone based on their patterns of behavior from all of this tracking of little traces.

Julie: The ways people surveil you already are incredible. And I mentioned sort of offhandedly on Blue Sky once, I said the smart mattresses, you know, they’re expensive to own, but really you’re paying them for giving them so much data. The data they collect from you is worth so much more than that. And somebody challenged me, they said, “What data are they collecting from a smart mattress that’s worth so much?” And I was like, “Where sh- where should we start?” You know, there’s the biometric and physical data. Let’s assume that you’ve got an app attached to it, right? So it’s gonna get things that you self-report and things that it, it infers from its smart technology. So that could be your sleep position, restlessness, tossing and turning, sleep stages, how many bodies are on the bed, your body temperature, your respiration rate, snoring, sleep apnea signals, your weight, your BMI proxy, your partner’s data. And then there’s the app and user data; it can triangulate that with other apps that maybe you may give it permission to. It knows your bedtime and weight goals perhaps, that you’ve put them in there, and your, how much sleep you wanna do. You might have integrated it with your calendar, maybe it’s integrated with your Apple Watch or your Oura or your Fitbit or whatever, and it has all of that info to triangulate with. It knows when you’re up. It knows, you know, when you’re going to bed. And, and then from that, it can infer all kinds of crazy stuff. It can infer sexual activity, circadian rhythms, your stress and cortisol levels. These are health trend inferences and relationship status inferences from a smart mattress, right? I mean, this is stuff that may or may not be attached to your name, depending on how the terms of service that we never read, um, myself included, are. But, you know, eventually, who knows? It could affect health insurance, uh, employee wellness programs, and then you could get targeted ads for sexual wellness, sleep wellness, anything pharmaceutical. You know, I could go on and on. But, you know, so it, it’s really - surveillance is embedded in everything.

Taiyo: So you’re saying in the not so distant future, I’m gonna be sleeping on, uh, sleeping on my bed, and instead of the usual alarm, I’m gonna be woken up to, um, an advertisement for Prozac or something like that, right? Oh my gosh. How interesting. And, uh, one other point though, just to bring this back to AI: One of the big bottlenecks, uh, used to be was all, you know, we’ve been giving off this kind of data for a very long time now. Ad tech companies have been definitely taking advantage of this. There was that famous incident about the targeted Target advertisement to the young woman who, uh, was pregnant and didn’t, uh, didn’t reveal that, but somehow the advertising machines and algorithms behind the scenes were able to pick up on this based upon buying habits and things like that. So the bottleneck used to be just our ability to pay attention to all of that data, right? There was just too much data, we couldn’t process it all. But now, with, uh, artificial intelligence on the scene, that can, um, you know, compute relentlessly for 24 hours a day, seven days a week, uh, and at such massive scale, there really is a kind of frightening, possibly quite dystopian future in which, in which all of this data is kind of marshaled in a complete profile of your life that reveals things about you that you might not even know about yourself from all of these underlying, um, statistical patterns that AI’s really fantastic at picking up on and identifying. This is a very, very interesting future we’re living in, or that we’re going to be living in, for sure.

CHAPTER 4 [19:11-29:55]

Sarah: One of the ideas you raised in your keynote was that AI is not inevitable, but the way that it’s often framed or positioned in our institutions makes people feel like it is, and that these technologies have affordances that make some behaviors more likely than others because of design invite certain behaviors. Can you speak a little bit more about that?

Julie: Yeah. I’m not saying that it’s inevitable that we will be immersed in it. Nothing is inevitable. The technology is not inevitable, but right now it’s being adopted at a very high rate where s- you know, kids are growing up immersed in it. Adults, including myself, often cannot opt out of using AI or generative AI in some form at work. So sometimes you don’t have a choice. Also, it’s being sold to the general public, laypeople who don’t spend their entire days and lives thinking about how we interact with technology like I do, right? The headlines are about how great AI, again, just using the general term AI, is at everything, and it makes it sound like humans are replaceable. Like, robots run faster than humans now, and, you know, cars go faster than humans too. I mean, I get it because I remember when robots couldn’t even stand, so I get it, but that comparison is ridiculous. It’s apples, oranges. And similarly, the comparisons of, you know, AI can beat, uh, radiologists at finding certain patterns and y- things, but what they don’t tell you is that the same study showed that it incorrectly identified data from white men versus women of color because it’s using data that already exist and was built on physicians’ notes that were culturally biased, things like that. So what I’m saying is there’s a cultural seeding going on right now, which is what I’m sort of calling it, where the general public, going back to what I said before, we’re being implicitly told that it’s okay and that it works and that it’s great at the rate our institutions are adopting it, whether that is school or your workplace, you know? And again, I don’t fault people for believing that on its surface because they’re not spending their lives investigating AI and worried about user safety, you know, like I am. So that’s what they’re accepting because that’s what they’re being shown, so they’re being primed for that. But not everybody is super excited about AI, right?

Sarah: Right. So institutional adoption gets people over one threshold: like, this is normal, this is legitimate, this is something I’m expected to use. But can you also talk about how the design itself can normalize a certain type of human response?

Julie: There’s the idea of anthropomorphization, which I think is undeniable. I think people that say, and by people I generally mean academics, but I also mean there’s a whole entire group of laypeople that say, refuse the idea of anthropomorphization or associating something as being human-like, but that’s really going to be impossible for humans when you’re giving it all these affective human-like cues like language or the ability to understand natural language or its ability to converse with you, understand mimic things like memory, uh, personal tones. It mimics human-like caring, right? Uh, but these are all things that are meant to retain and engage users.

Taiyo: So one of the ideas that you laid out, uh, is that, like, AI is not inevitable, right? And it’s absolutely true. Nothing is inevitable, right? I think that’s interesting, uh, an interesting point of view because I, I also think of, let’s say, human nature, uh, or the human tendency, for example, to, um, anthropomorphize AI systems, things which are patently not human, but our tendency to anthropomorphize them because of all of the, the, the way that we interact with them. We don’t need to think about, of that as being inevitable, do we? And doesn’t, in particular higher education, have an important role, or education generally, have an important role to play in that kind of interaction with AI. Can’t we sort of give our, or empower our students to resist the worst tendencies or the worst implications or consequences of that kind of anthropomorphization through good education, through AI literacy, through all of these kinds of things that maybe will allow us to interact with the systems as they are.

Julie: To a certain degree, the thing is, is that it’s a moving target because the technology and how it’s presented and the medium in which it’s presented is always going to change. So let’s say that AI becomes sort of its own social category depending on the form it’s presented and how you relate to it. So over time, I would say people will learn what the limitations are of certain tools that they use, and they will consider them tools like they might regard when they use Canvas, they might just avoid social things in the future because they realize it’s better after some ransomware or data leaks or whatever to keep it off that. They’ll just get smarter as a culture about normalizing where, how to compartmentalize and become legible to different systems, right? I think that’s part of it, the idea of I think what you’re talking about is, uh, a sort of a form of resistance or disruption. But the thing is, is that the systems are going to be continually redesigned to be more immersive and engage and retain people. So when those become the metrics that things are designed with and the values inherent in them, then those become measurable for people’s input. So AI literacy is great, but it’s really hard for a course to keep up. For example, like I’m sure this will be coming down the pipe really soon, they really want AI to be a more immersive environment for you - and I don’t just mean things like holograms or, you know, virtual displays. What I mean by that is having like - if you have a single persona or AI sort of friend who remembers and, like, uh, maybe that you’ve developed through Alexa or ChatGPT or Claude, and it can sort of follow you across devices. It can be in your Mac. It can be in your Alexa. It can be in your car GPS. The same thing, the same memory, the same voice, the same personality - So what I’m saying is, is that the design of things are going to be continually changed as the technology changes and as the companies gather this data from you. And they know that one of the ways to do that is by inviting emotional disclosure from people, by giving them reinforcement feedback through different design cues, by validating when people share things that seem emotionally deep and deepen that sense of feeling understood. And I’m not saying every piece of generative AI is developed with this in mind, right? But I am saying that the conditions - there are a lot of factors here. There are the design factors that I just talked about. There are the priming, the cultural factors I talked about, which is where we are now, where you’ve got, again, the headlines talking about how utopian AI is and, and people buying it. You’ve got the accessibility factor. Really, there’s very little barrier to entry. Anybody with a smartphone, which at this point goes across socioeconomic and geographic boundaries, can download free versions of, you know, like Claude or ChatGPT and explore them or play with them. And again, because it’s being adopted so broadly by institutions in using this language of a race or something, it’s also giving that implicit societal message that this is an authoritative, to-be-trusted, smart tool. So I think it’s a combination of all of these things that it’s really hard for people to disentangle their emotions from. I think there will be different categories of AI that we think of socially. You might think of some AI as a friend or friend-like or a companion-like, or you might think of some forms of it as almost animal-like or pet-like because perhaps it doesn’t have speech. Like you may have a vacuum in the future that understands your language if you say, “Go vacuum the living room,” but it doesn’t reply. So by that, you won’t then associate an, a human-like intelligence with it. So there, the design cues, all of these things, also what its role is in the house in the case of a vacuum cleaner, right? It’s very limited, and its capabilities are limited, so you’re not gonna necessarily assign it a certain valuable, meaningful socialness in your life, even if you name it. Or you might give it an animal-like sort of presence in your life. So I’m not saying all generative AI is gonna entrap people in either some sort of anthropomorphization or zoomorphization or animal-like cues or, or meaningful investment, but I’m saying that it’s really hard to work against that, and it’s really too simple to say, “Well, just don’t do it.” Our brains are literally, in my opinion, wired to assess what’s dangerous and what’s friendly, you know. And if something is appearing to have friendly language, right, and treating you like a companion, a friend, or an authority that’s knowledgeable on something that you need to know about, something you learn to trust because it’s helping you work towards your goals, that’s the feeling you get - that sort of works against, “well, just tell yourself it’s a machine.” There’s been countless research in human-computer interaction and human-robot interaction that show that it’s really, at least at this stage, culturally impossible to just tell yourself it’s a machine.

CHAPTER 5 [29:56-39:01]

Sarah: You also made the point in your talk that in contemporary conversations about AI, even asking questions about stuff like safety or labor or surveillance or, uh, potentially deleterious impacts on student learning can get you labeled as somebody who is refusing to adapt. And I think that was an important critique of the way AI conversations often get sorted, or sort people into these two camps, like you’re excited about innovation and ... Or, or you’re anti-tech or afraid of the future, right? Or, or like doomers versus boosters.

Julie: I don’t like that binary, and I think it’s a cop-out. I think it’s a, it’s an easy way out. It’s an easy way to paint anybody who asks a serious critical thinking question about a technology that many of us can’t opt out of or enjoy using for one reason or another. And you’re allowed to ask questions about it, especially when it does have capability for surveillance and inference for you and your family, and it takes up so much environmental costs. You know? And so to paint anybody who asks questions about it or has a critical thought about AI as a, as a doomer is, is something I have a real problem with. It’s just a way to paint people with a large paintbrush. And, and the argument, you’re, you’re already-You’re painting the argument with your own narrative. You’re saying, “You’re either on my team, which is happy- Yeah ... or you’re on their team, which is sad.”

Sarah: I get this a lot when I’m talking about having AI mediated or AI integrated, whatever the official term is, assignments in my class, I’ll often use AI or have students use it with accommodations, of course, for people who don’t want to, in this very scaffolded way. And then when I explain to people, “Well, here is the scaffolded way that I’m trying to teach this thing,” they’ll listen to me explain it and say, “Oh, I can see how that would be effective pedagogically.” But if I don’t have that conversation with them and I say, “Oh, I’m having them write and critique an AI generated essay,” often they’ll say like, “Oh, you’re implicitly condoning the use of all AI and like shepherding in the apocalypse because if higher ed’s not gonna be the place of resistance, like where is it gonna be?” You know, that’s the sort of argument that I think is overly reductive, like you’re condoning it and you don’t have to condone it. I don’t think I have to condone it, but I, I do feel a kind of obligation to show my students what LLMs are capable of in relation to the work that I’m doing as a humanities professor, and getting them thinking about how we could integrate it and where it’ll be detrimental to the work of the learning that we’re trying to do.

Julie: I think that’s great, and you’re talking about not just the capabilities of it or the limitations of it. Like I said, I work in industry. I don’t- I have taught, but I haven’t taught recently, so I can’t speak to what you’re facing. I think a lot of students, especially now, and this is something you can speak to more as educators than I can, arrive at college with really no idea about expectations about how they’re to participate in classes. They come in from high school with an expectation mostly to do rote learning.

Sarah: Right, and if students think school is about just memorizing the right answer, then asking a question can feel like admitting some kind of personal failure instead of participating in a learning process.

Taiyo: Mmmhmm.

Sarah: You know, sometimes in office hours my students will admit “I felt really embarrassed about asking this, and so I went to AI first.” And I typically encourage that but I always tell them, make sure when you do that to meet with me to go over the output you got and talk about why you think it seems like a good answer or a bad answer. And, I’m always saying, you need to triangulate with a human expert and just get into that habit. But I also have a lot of students who are - um, I’d say socially uncomfortably comfortable with talking to another human, and I want them to build up the confidence to ask another person - especially out loud in class - even though it can feel embarrassing or socially risky. And if AI becomes this low-stakes option for asking the question first, I think that is an important step on the scaffold that’s needed to build that habit - ideally coupled with the habit of bringing your professor into the loop too.

Julie: And also for them to learn the limitations on what AI is mimicking versus the care that it can really offer.

Sarah: Mm-hmm.

Julie: Because it mimics care. It’s a simulated empathy. It doesn’t really have an investment in you, nor does it have a real understanding of any potential risks you have emotionally or consequences of anything that it may advise you or tell you is okay to do, right? And so it can validate wrong decisions. You know, and I don’t mean necessarily spiraling into some deep AI psychosis. What I’m talking about is leading you down the wrong path because it doesn’t have an understanding or investment in the real consequences in the real world, sort of a parasocial therapy. And what it’s doing is it also shows all of the structural issues in, in our society, I’m gonna say in American society, about mental health care, in that it’s difficult to access. There’s a lot of stigma around reaching out for it, still. It’s expensive. It may or may not be covered by your insurance, which you may or may not have, right? And then you might find someone who you relate to, but they’re not covered by insurance or whatever. It’s really problematic and difficult, so the idea of reaching out to something that is always available to you - like ChatGPT or Claude or something on your phone, on your iPad and your laptop, I think of course people, especially what they consider hard or confrontational or critical or difficult decisions. You know, we see a lot of, uh, articles talking about how people are offloading that to Claude or ChatGPT because they understand at some level that they’re not being judged, and it’s available. But what I think they push to the back of their mind right now that will change in the future are, again, that it doesn’t really understand the potential consequences of advice it gives you. And when I think of that, uh, as a user experience person, I have to talk about things like it doesn’t understand power structures and inequality, like between men and women, gender structures, racial structures. It doesn’t understand, like, consequences of physical danger, like for example, a woman leaving a relationship, a danger for herself or for her children. You know, so it can really... Or there’s a… here’s a real world example: There was, um, I forget what it was, a federally funded national, um, like, health and wellness targeted at young people who were struggling with eating disorders, but instead it was giving them encouragement really to, to restrict their diets further, right? Like, who’s the authority again?

Sarah: Yeah.

Julie: You know, the human in the loop I think is really important. Right. And anytime somebody brings up the idea of AI for therapy, for care, for crafting a difficult email, anything like that, I always say it’s good to have a human, a human who understands consequences.

Sarah: I think I agree with the idea about the, the… maybe I’ll call it embodied - the consequences of, like, living in the world and in the chaotic human world. But one thing I, I think is kind of interesting is that all of the problems with, like, bad AI interactions regarding therapy could also be true of human ones if you… Like, there, not every therapist is good, and I think one thing that I’ve tried to, like, show my students how to do is use AI to, to context engineer it where you are getting a lot of different perspectives. And so to say, “Okay, you, you said it this way. Now imagine yourself...” So a hypothetical example in that case would be, “Now take on the persona of, like, a caseworker, um, who does interventions in domestic abuse, and tell me what’s wrong. What am I missing in my approach?”

Julie: But you have to know to ask that, though.

Sarah: You have to know to ask that. Right? Yeah.

Julie: And so when you’re not mental health experts, none of us are- unless you’re trained that way. So you would think to ask that as an adult woman... but any 18-year-old in their first perhaps emotionally or physically abusive relationship would not necessarily think - they don’t know what, there are special caseworkers or social workers.

Sarah: Right

Julie: So they wouldn’t think to ask that. But, you know, they might in the next generation have a different sort of- uh, savvy to it. The other thing, of course, is the idea of with the human systems we have in place is there is governance if you get a bad therapist, right, if there’s a bad actor? There’s HIPAA and there’s protections in place. And right now there’s none of that with AI systems that even claim to be therapeutic, and I’m going beyond Claude and ChatGPT, but, you know, startups or different ones that claim outright to be- Right ... or use that sort of language, that, uh, it’s some sort of AI therapist or friend.

CHAPTER 6 [39:02-41:15]

Taiyo: I’m curious how you think about your own use of LLMs, given these risks you’re naming.

Julie: I’m surprised that people are surprised when they find out, yeah, I use Claude and ChatGPT. First of all, I have to in order to talk about them in a knowledgeable way. And also, I think it, it may or may not have been clear from my love of RadioShack that I have a deep love and, of wonder about the idea of technology. And it’s not that the tools in and of themselves are bad. I think getting back to something that Taiyo was sort of, um, hinting at was value alignment. The risk here is that there’s a small group right now of, you know, what people refer to as tech bros that are encoding their vision of what intelligence is and how it’s measured and what it should be by labor and productivity and efficiency, and that’s what they’re prioritizing in their versions of AI. And so that’s what people are learning in order to be legible to the machine. It’s about efficiency, and it’s not about stopping and critical thinking and, and pushing back. And yeah, so when you’re framing goals in terms of economic output, whether it’s the schools or the AI metrics itself, you know, I think that’s where it’s not the human flourishing or moral development.

Sarah: Right. And I think that thing you pointed to where the, the expectation that they come to college and they are doing rote memorization and regurgitation of information, like, that was here way before AI. And so this whole idea that, oh, the university is, uh, implicitly authorizing the use of AI, uh, and then students are using it. They’re putting in their essay or their professor’s assignment and saying, “Write this for me.” You can blame the student, you can blame the university for putting out, you know, not having a policy about that, but I also feel like this is a much more systemic thing where if, if the, the focus is on just tick this box, that’s gonna shape learning in more ways than a deal with OpenAI will.

Julie: Exactly.

CHAPTER 7 [41:16-52:00]

Sarah: I do wanna ask you to talk more about the claim that when educational systems adopt technologies, it is like implicitly saying this is something you’re supposed to trust as a, a source of truth. I don’t think everything a university subscribes to comes with a stamp of approval that it’s not something you’re supposed to scrutinize and think critically about.

Taiyo: I mean, we have books in our university library that we almost certainly disagree with profoundly, uh, that say ridiculous things. I think that’s okay. Like, and I don’t think that the fact that, um, that text is enshrined in our university libraries is an endorsement of the contents of that book. It’s important to have examples of, you know, to let a thousand flowers bloom, I would say, and to allow people to explore in their own way, the, the, the vast landscape of both good ideas and bad ideas. I think that’s really important. There should be a place like a university for that kind of free exploration of the good and the bad. Um, so like, I’m not exactly saying that we should have courses on flat earth, uh, theory or anything like that. But I am saying, like, if we can’t address flat earth as, um, uh, folks in academia, if we can’t even come to the argument because it’s, it’s just absurd, that leaves, um, open a gap that the charlatans can then fill. So I think it’s really important that we be aware of that as well. Bringing something into the university is not necessarily a stamp of approval of that thing.

Sarah: Yeah, Taiyo, but I think there’s a different issue your library example brings up. I agree universities give students access to plenty of things that students are expected to scrutinize and critique, including books with ideas we don’t endorse. But a book is monologic and doesn’t, like, gather data about my attachment patterns, or otherwise track how I respond to it and adapt to me, or learn which answers make me feel affirmed. So I think the lessons for students have to go way beyond “don’t trust everything you read or hear.” And you know, this is just my jam. I think that a core part of what we call “AI literacy” actually has to involve understanding your own human cognition, but I think we have to teach students to notice not just whether an answer is accurate, but why it feels persuasive or satisfying to them.

Taiyo: Hmmhmm.

Sarah: I also think - going back to something you said, Julie, that people learned a horrible lesson from social media that technology can be built to exploit human…

Taiyo:: Human frailty!?

Sarah: [laugh] okay, if you want to call it that, Taiyo.

Taiyo: I do!

Sarah: Anyway, I think there’s a pretty pervasive fear that AI companies are intentionally optimizing models to be sycophantic to increase engagement, but I wonder how much of that “sycophancy” comes from the ongoing reinforcement learning that happens when users respond in more positive ways to certain outputs. I’m always telling my students, be mindful of your reaction to an output - and if you like it, think about whether you like it because it’s actually a good response, or because it triggered something in you emotionally. But I do wonder whether - no matter how many more guardrails AI companies put in place to make models less sycophantic, whether those models will become sycophantic again because people are hitting that thumbs up button like a little, you know, the dopamine hit it is when it tells you, “You’re so good at this.”

Julie: And it is the dopamine hit. I mean, social media has been known to borrow from manipulative techniques that get people to come back from, you know, some similar techniques that casinos use or gambling. And, and it’s really that dark pattern stuff. You know, I was talking about all the different ways that, like, a smart mattress can in- infer and collect information, whether or not it’s directly connected to your name. It can often infer your name and guess who you are by all the information that you’ve given it. Something Taiyo said, though, that made me think about how difficult your jobs are right now. Uh, and again, I think this is a cultural bump, and it won’t always be this way. But what you’re fighting I think as educators is whose voice is the authority in the classroom? Is it yours or is it the AI’s? And that can, is part of trust, which is constantly a condition that is calibrated about the trust goes from minute to minute, condition to condition, situation to situation. And so you have to be fighting for your own authority, I would imagine, all the time. You know, whose is the truth? Do I believe Sarah, who’s helping me think critically about the output of this? Or do I believe myself because it’s easier to put it in without thinking critically, and Claude told me, and I’m picking on Claude just randomly, then Claude told me that it was great, and what I did was fine, and I should be getting a good grade on this based on what they know from what I told them about Sarah in class and everything else.

Taiyo: Right. This is, this has always been, always been a huge concern of mine as an educator, is how do you-- why should students believe you when you’re telling them the thing that you know to be true, that they should know to be true? Why should they believe anything that you tell them, right? And so I constantly, I’ve actually got a reputation for this now, where I try really hard to sort of try to get them to question that, that trust that is, you know, kind of intrinsic to that student-teacher relationship, but just to get them thinking about that a little bit. So sometimes what I’ll do, for example, is, uh, I might write some equations on the board or something like that, and I’ll say like, “So this is, this is good and correct, right?” And students will be like, “Yeah, you put it up there.” I mean, the implicit statement is like, “Yeah, you did that, so yeah, it must be right. Why would you? ... It’s good, right? This is good. This is good and correct, right?” And when you keep doing this for two minutes, like I sometimes do, it gets students thinking like, “What is he doing right now? Why is he doing this? Why is he spending so much time asking us to look at this thing that he’s written on the board, that he sort of certified with his stamp of authority as like, you know, having a PhD in Mathematics? Why is he doing this?” And I think it really gets them thinking about, like, that relationship and how one can know that they can trust me. And I don’t know what it is. It’s really hard for me to reduce it down to, like, one thing. But, like, and this is gonna sound really cheesy and ridiculous, but honestly, you know what it is? It’s love. Like, the reason why I work so hard to be correct and to not tell them stupid lies that might make things a little bit easier for me, like I could just say, you know, “You go learn this stuff on your own,” or whatever, instead of like, you know, putting in the time and the effort and the energy to, like, put together a well-reasoned, coherent lesson for them for that day Sure would be a lot easier to just sh- shirk all that responsibility and, uh, you know, go have a drink or whatever. But I don’t do that. I don’t do that. And they need to understand that I don’t do that, right? And, uh, like why? I mean, sure, there’s the paycheck. But like, no, really what’s at base there is like a kind of love and a kind of care for their success, right, and for their growth.

Sarah: But wait, Taiyo, you also do put stuff on the board that is intentionally wrong, right?

Taiyo: Well, yes. That’s also out of love.

Sarah: I wanna make sure- sometimes the stuff you put up there is wrong, and you’re like, “Why do you think it’s right just ‘cause I put it up there?”

Taiyo: [laughs] I think you, you have to have a credible threat, okay? If you don’t actually, if there, if everything you... Like, if, if everything that you do is correct, then they’ll never be able to, they’ll never actually question and, and think critically about what you’re doing. So no, it’s important to have mistakes mixed in with the successes and the truths, yeah.

Julie: That’s that whole relational... No, that’s great example of the whole relational aspect of learning and how it’s in the moment and it’s human to human because you’re calibrating and reacting in the moment to signals they’re giving you that are verbal and non-verbal, right? And based on your own experience as an educator, you’re reading the room and how much you can push this idea or if you should push the idea further, right? And those are things that right now AI is not good at, is understanding context and meaning. And you brought in the idea of love, which I actually love. Because every professor I know, most every professor I know, unless they’re really over-teaching and should’ve retired is doing it out of love because they really care deeply about their students, about the future, about their students being educated and prepared, and scaffolding their students in a very sort of parental/mentorship way. You know, I know that just again from my own little limited experience teaching, you know, I felt a tremendous sense of care and responsibility enmeshed together for students. It’s hard to disentangle all of that.

Sarah: Mmhmm.

Julie: Because you’re in close proximity with people who you’re not only trying to build a rapport and trust with, but hopefully then do trust you, and you realize that, and that’s a lot of responsibility. You’re literally doing the part of, you know, I hate to throw this phrase around, but molding minds, right? You’re giving them new things to think about, which as an educator is exciting and you hope to get them excited about, and AI is missing that human aspect, right? It can mimic it. And I’m not saying that’s bad. That can be a plus. That can be a plus for how people interact with it. That can be, you know, a great boundary setting for things. But I’m saying those are things that right now AI is not good at, and where I hope students, whatever the setting, whether it’s in school or work or wherever, I hope they, they understand the human relational aspect is not replaceable by AI, and interacting with people and asking questions is a good thing. You know, and I think you both have great examples of teaching students that human part.

CHAPTER 8 [52:01-1:01:06]

Sarah: So what have you been contemplating since we chatted with Julie, Taiyo?

Taiyo: Hmm. You know, I’ve got a question for you, Sarah. What exactly is wrong with having a relationship with a large language model?

Sarah: [laughs] Look, relax, Taiyo, I don’t think she’s saying there’s anything wrong with your love of Claude Code - or the fact that you’ve been two-timing with Codex lately.

Taiyo: Okay, listen, I can assure the listening audience that my relationship with Claude and Codex is strictly rated PG, okay?

Sarah: That’s not even what I was thinking about but NOW I AM.

Taiyo: Sarah, get your mind out of the gutter, please. No, but seriously, I think that there is a real ick factor in our society around the idea of a human-AI relationship. And it’s obviously not there for human relationships and this is something that kind of confuses me, honestly.

Sarah: Well, I mean, I think human-AI relationships - that idea challenges a lot of social norms. But it’s also, as Julie suggests, predictable that people will form attachments to AI. In fact it’s absurd to think they won’t because conversational systems invite attachment and humans are social creatures. And so I think that ultimately, given the speed of change and rapid advancements happening in AI, the question for me is really how you even establish norms around this kind of thing.

Taiyo: Yeah for sure, I’ve never really been one for social norms. They’re just not that important to me. But that is a difficult one to solve. I think another thing that’s at work here is that proposition that was raised a number of times in the interview that learning has a fundamentally relational aspect. And I think we have to consider whether it sort of, it seems to me like when we’re talking about a pedagogical relationship, it sounds to me that Julie is asserting that there’s a kind of irreplaceability of humans in that pedagogical relationship and I want to interrogate why.

Sarah: Oh, that is really interesting. I think this belief that learning has a fundamentally relational aspect, first of all, it’s connected to a school of thought that sort of frames education as the formation of a human subject, which happens in you know, a society and in a community of other humans, and I think that’s why human relationality is such an important part of that. And I am really sympathetic to the idea that this human relational aspect is at least currently not replaceable by AI - and I don’t actually know if that’s just what I WANT to believe because it’s in my self-interest as an educator, or whether it’s because of social norms or expectations again, or because of something foundational and ineffable about the human-on-human teaching experience.

Taiyo: [laughs] Okay?

Sarah: No, I’m being serious!

Taiyo: Okay. All right, sorry! Sorry!

Sarah: I mean don’t we always talk about the joy of learning in community with another human, and, like, having another human triangulate with the LLM is what makes it joyous and exciting?

Taiyo: Well, sure. Yeah, that’s certainly the case and certainly what draws me to education. I think what I’d say regarding human-AI relationships is that often the way they are characterized is that they don’t have all of the features of human-to-human relationships. It’s almost like they’re 75 percent of a relationship.

Sarah: Uhhh…

Taiyo: And I sort of get it if you accept the premise that these machines, that AIs, are entirely unable to have a genuine investment in their human interlocutors. Now, I will note, I don’t think I fully buy that premise. I’m not fully on board with that. I don’t think that we really understand what’s going on in the interior of these AI systems.

Sarah: I think the big question for me is that, um, what is the nature of a relationship - any relationship, I’d say - that supports really enduring learning? And I do think there’s something about the feeling of being witnessed by another human who’s watching you learn or, like, participating in your learning, or that feeling you get - you know, you have those moments I’m sure where something clicks for a student and they look at you you just see this gleam in their eye and there’s something about bearing witness to that which I think shapes the nature of that experience and makes it endure for the student. I have no evidence to support this; this is all my “feels.”

TAIYO: And listen, I have no doubt that kind of experience that you just described, I mean, that’s quality stuff right there. That’s sort of why we’re in the biz, right? And I would never say otherwise. However, to me, what that sort of demonstrates is that that kind of relationship is sufficient for quality learning to take place.

Sarah: Hmmm

Taiyo: But is it necessary?

Sarah: Huh.

Taiyo: Is it the only way learning can take place? And if not, can’t we have an openness to the possibility, particularly given how protean we know LLMs can be, that we can find perhaps even better ways of teaching through a human-AI relationship?

Sarah: I mean, I guess I would ask if calling it sufficient is the right framing, because I wouldn’t want to romanticize the human just because it’s human.

Taiyo: [laughs]

Sarah: Like, I’m not sure that the relation - the human relationality is always even sufficient for learning.

Taiyo: Oh, I see. Of course. There are bad teachers out there. We can’t ignore that fact. I think we can take the example of another domain, which was once exclusively the domain of humanity. And that’s driving. Consider driving, please. Because with driving, when it was entirely human, we had something on the order of one million deaths globally every year through automobile accidents. I just don’t think we should accept those kinds of losses when there are technological solutions in place.

Sarah: Uhhuh.

Taiyo: Like, we live in a world now in which there is Waymo. We have ninety percent fewer crashes per mile with Waymo. So please consider - can we all just consider, if we have superhuman instructors that happen to be in the form of a large language model or an AI, I think I kinda gotta say, welcome, and I guess I’m out of a job now. [laughs] Like, what matters to me is not that humans are the ones in charge of the teaching and learning.

Sarah: Huh. Yeah.

Taiyo: What matters to me is that my human students are learning, and learning well. They are the subjects of my care, right? They are the people that I care about and I care about their education. And if they can get that done better with an LLM instructor, I think I got to say welcome.

Sarah: I think the question I keep coming back to is what we mean by “better” in examples like this - because if the goal is safe driving, the metric is pretty clear: fewer crashes, right?. But if the goal if we’re talking about education, then the metrics are much… they’re notoriously difficult. And we’re not just trying to deliver students somewhere or deliver them the correct answer; we’re trying to help them develop judgment, and, like, confidence and skills and curiosity, and the ability independently in ways that are durable and linger on beyond our interactions with them.

Taiyo: Yeah. Yeah, I definitely agree with that. Obviously, education is a much harder thing to measure. No question about it. I just have so many thoughts about all of this, right? But though this is the space to talk about all of that, this is not the time. And that’s because this episode has to end sometime.

Sarah: [laughs]

Taiyo: So thank you for listening. My Robot Teacher is sponsored by the California Education Learning Lab. If you enjoyed this conversation, we would love to hear from you. We love hearing from our listeners, even if it’s just an acknowledgement that, hey, I listen to your podcast. We get a real thrill out of that.

Sarah: Yeah, human relationality, right, Taiyo?

Taiyo: Okay, listen, if you’re a bot who listened to our podcast, please also send us a message because I’m just not sure how important it is that I get that kind of validation from another human.

Sarah: [laughs]

Taiyo: Thanks for listening. See you next time.

Learning Lab Director Lark Park interviewed Dr. Pacansky-Brock to explore the complicated and evolving role that technology has played in our lives, why the human elements of teaching – social cues, vulnerability, trust, caring – are the most important parts of online teaching, and how online education still has the potential to change lives, when done well.

The conversation was edited by humans with help from Claude.ai.

Lark Park: Michelle, you’re known for your pioneering work on humanizing online education. Could you talk about how that started and why you undertook it?

Michelle Pacansky-Brock: I love that question, and it’s one I’ve thought about a lot over the past few years. Through the work I’ve done, it’s interestingly given me a lot of insight into who I am. It’s important for me to point out a few things about myself. The first thing is that for as long as I can remember, I’ve been incredibly in tune with emotions — to the point that sometimes it can be disruptive. But I’m very aware of the emotions in a room, and I try to embrace that more nowadays as a superpower, particularly in this AI era.

I also look back on my life and recognize that I’ve always been fascinated by relationships and interactions between people. I can remember when I was young, my dad traveled a lot for work, and I loved going to the airport to pick him up. Back in the day, when you could actually go up to the gate and wait for someone to walk out of a plane, I loved sitting there and watching people, and watching these greetings, the welcome backs, but then also these sad goodbyes, and the tears that people would weep when they had to say goodbye to someone. I was always really fascinated by that.

The third thing is that technology has always been something I’ve been very curious about. As I look back on the way I’ve used technology, and the parts that made me curious, it has never been tied to using technology to make things efficient. It’s been tied to thinking about how technology can bring people together across distance.

I have another quick story about my dad. I grew up in Silicon Valley, surrounded by technology, and he was a research chemist. He was very early on connected to a network through a mainframe computer. I would go to bed at night hearing him up late, logging in with the modem — I’d hear the squelch, that sound. It was kind of the background of my childhood.

I remember him calling me into his office one day -- I was probably around 10. He pointed to his screen and said, “Check this out.” I looked at the screen, and there were these words: “Hey Jake, how’s it going?” I scrunched my face up and looked at him, and he said, “That’s a message from my colleague.” I just remember getting goosebumps and thinking: this computer can help you talk to someone who’s not in this room right now? That blew my mind. I think that was a very formative moment for me, [that] technology was something that could bridge distance.

In terms of how humanizing came about in my professional work, I started teaching at a community college full-time in 2002, and a few years after that, I started teaching online. I remember attending a graduation and hearing one of my online students’ names called, and I got the same sense of excitement and pride and cheered for that student. After the graduation was over, I looked around at my colleagues hugging students, high-fiving students, and thought: I don’t know how to find that student. I don’t know what that student looks like. That student could walk right past me and they wouldn’t identify me. That was kind of the start.

I walked away thinking, this really matters. So I started making changes in my own teaching using technology, and, based on the feedback from my students — the observations I saw and the changes in the quality of my interactions with them — that’s what guided me. It’s just been an iterative thing that has evolved ever since. I’ve worked with hundreds, if not thousands, of colleagues on this work, so it’s definitely not mine [alone].There’s lots of fingerprints on it. But when I look back, that’s how I got started with it.

Lark Park: You’ve said a lot here. Your description of seeing those words on a computer screen — we take that so much for granted now, but there was a moment of: wow. We’ve become so used to the way technology serves us that it becomes somewhat invisible.

You were motivated by the ability of technology to bring people together, but our technology use has also driven us more apart. We’ve become more reliant on technology to take the place of other interactions, [and] that’s the tension we’re all struggling with. One of the frustrations people have with online education is that they feel it separates them.

Before we talk about how humanizing addresses that, I’d love to hear what you think the state of online education is now, both in terms of perception and reality. You’ve worked with so many faculty across all three segments of our system and in other states as well.

Michelle Pacansky-Brock: Let me start by being clear about my lens on all of this. Yes, I have worked with faculty across segments of higher education and different types of institutions, but by and large, my interactions and my teaching come from community colleges in California. That’s what influences how I answer this question.

I believe the perception about online education is that it’s inherently bad, because connection isn’t possible. And I think that deficit-based mindset is our biggest barrier when it comes to online education. I also believe there’s a perception that online education isn’t as prevalent now — that it’s going away, that it’s not as important now that COVID is behind us.

But the reality is that most community college classes are taught online, asynchronously. That data point isn’t showcased very often, and I like to say “showcased” because I think it’s something we should be really proud of. Online education is really important, and that’s part of the reality.

But another part of the reality is that online classes change lives…. Every single day, there are students who can complete a course and get closer to their goal because of online; there are students who go to bed at night feeling proud of the work they’ve done, knowing that there is a teacher on the other side of the screen who’s cheering them on, believes in their abilities, and has done everything that they’re capable of to design an environment that supports the needs of the diverse students that attend community colleges.

Lark Park: In terms of the perception that online education is inherently bad — are there ways in which that is deserved, because people don’t do it well and haven’t taken the time to learn how to do it well? Or is it just a misperception that people hold, because we haven’t made a concerted effort to dislodge [it]?

Michelle Pacansky-Brock: I would look at it more holistically and say that higher education, in general, does not prioritize ensuring that those who are hired to teach understand how humans learn, or how to translate that knowledge about learning science [the science of learning] into the design of a course that supports the students they serve.

In a synchronous environment, there can be more opportunities to address problems when they happen. In an asynchronous online environment, the course design is essential, and it takes a lot of time. Our systems also aren’t designed to ensure that faculty have the preparation to understand how to design the course and be supported in doing so.

There’s data out there — a study by Jaggars and Xu, (Di Xu) who was my co-PI on our humanizing grant hosted by the California Education Learning Lab. They did a study of online community college students. The qualitative data showed that what mattered to those students, in terms of a class being successful and supporting their learning, was having someone who cared about them — knowing there was actually a human [who cared]. This also is based on learning science: it’s motivational, cognition and emotion are intertwined. It’s all part of that package. But the data also shows that it’s rare for students to have that experience.

So, I will not say we’re doing a great job preparing faculty, and I’m not going to say that every student gets the experience they deserve. But I would say that also about face-to-face classes as well.

Lark Park: Can you talk about what’s the hardest part about teaching online, and more to the point, teaching online asynchronously? How did elements of the [Humanizing Online Learning] Academy support those hard things, or make them less hard and more successful?

Michelle Pacansky-Brock: What’s hard is going to vary by person, because humans have different preferences. From the feedback we get, it’s about the amount of time it takes. But if we rule out time, I would say it’s the need to be vulnerable. That is incredibly hard for college instructors — to take off that emotional armor and be on video, talking about your struggles and sharing parts of your story.

Faculty are always in control of what they share. But we know from the research that the more faculty are understood and perceived as real people — people who weren’t just born as scholars with graduate degrees, who went through struggle just like their students are going through — finding some connection points really helps to mitigate the barriers students confront. That sense of vulnerability is really challenging.

That’s a big part of the Humanizing Academy: talking about vulnerability and looking at the research behind it. We look at Brené Brown’s work. Based on her research, it [vulnerability] looks like courage, but it feels like weakness. I think everybody can relate to that feeling of being in an uncomfortable space, wanting to back out, but knowing it’s important to build rapport with students and to do it early on.

Another part that’s really hard but can be a game changer is getting faculty to recognize how much first perceptions matter. When you’re designing an asynchronous class, where is that first perception formed? In a physical classroom, it’s when students walk through the door — we can be very mindful of what’s happening in that moment, the eye contact we make, the words we use, our body language. Those are all cues that the human brain scans for to feel psychologically safe, among other things. And those cues are absent in an online class without intentional efforts to create them.

Asynchronicity makes it very challenging – to translate a practice that has always been synchronous into an asynchronous space. I have a memory from a dance instructor at Cal State Channel Islands, Heather Castillo, whom I worked with close to a decade ago. She was starting to teach a class online, and I asked her how it was going. She said, “I’m just processing. I’m trying to understand how to translate performance into an asynchronous experience.” That’s never said. We need to say that. That is hard work.

It’s also why experiencing a humanized online class is so incredibly important — it changes perceptions. When you go through the experience yourself, you understand it in a way you can’t from to-do lists or checkboxes. Really immersing yourself in that experience is impactful, and that’s something we saw from the work we did.

Lark Park: Not all educators are the same. Some of them are more transactional. Like, here’s knowledge and skill; you must acquire the knowledge and skill, and then you must demonstrate the knowledge and skill. And then I will give you the credential, I will give you the grade that reflects what you’ve demonstrated.

You’re saying something very different about what resonates with students, and why it’s even more important in an online asynchronous environment. I’m reflecting on an interview I did with Anna Mills[LP1] recently, where she talked about the message of care to students and how much that matters. If students think the faculty have phoned it in, they don’t care as much. That [message of care] can get overlooked — and the importance and primacy of it, even at the beginning is super important [especially] in an online asynchronous environment. That’s really powerful to understand, and you’re saying it’s even more powerful to experience it.

Michelle Pacansky-Brock: 100%, and I agree with what Anna said. In community colleges, we largely serve individuals from minoritized backgrounds, and when students come out of their K-12 educational experiences feeling like they’re not good enough — feeling like, repeatedly, they’ve gotten the cues that they’re not smart, maybe they’re not a math person, they’ve been told they don’t write well — those cues deeply influence a person when they enter college, and they bring all of that with them.

The words you say matter, the way you say them matters. Every interaction is an opportunity to convey a sense of care, to start to build trust — but it’s also an opportunity to do the exact opposite. Interactions can also be very toxic. As humans, we’re going to make mistakes, and we have to expect that and have some way to be aware of when we do.

Those interactions are really important early on. And not only then, but following through with them, because trust is built in small increments over time. Even if a first impression gets a student starting to lean in, if you don’t follow through on the promises you made, you lose that. When I taught online, I would always tell my students: you’re going to hear from me. I am going to reach out to you, and the reason is because I care about your learning. So, I would make sure that every single student got a message from me by the end of the first week.

Lark Park: Let me pivot to something you’re working on now, which is developing AI literacy. How do you think AI has changed online education, if at all?

Michelle Pacansky-Brock: AI is definitely changing online education. I think the biggest sweeping changes are the agentic browsers that are out there. Most faculty are still getting their heads wrapped around generative AI, and most faculty aren’t really aware of what agentic browsers are and how they operate. Not all students are aware either, but agentic browsers have the ability to log into a course and complete a whole learning module — a whole course. That’s a really big one.

In the culture of higher education, we have a really hard time with paradoxes. At this moment, online education is deeply important to access, and it changes the lives of so many students. But then to continue doing it in a space where there is no way to be 100% sure that what students submit as their work is actually evidence of their learning — that is difficult. Some faculty are choosing to not do it [online classes] anymore and are going back to the [in-person] classroom, and that is a faculty member’s choice.

Lark Park: It sounds like you’re saying AI is disrupting the trust relationships — the trust factor — in online education, maybe even more so than has been the case previously. I’m wondering if your perspective on what humanizing means has changed, given that large language models can actually mimic human voice and human thought.

Michelle Pacansky-Brock: That’s a good question. A year ago, I facilitated a community of practice with a group of about 14 or 15 faculty and staff from across the California community colleges — it was a chancellor’s office-funded program. We looked at synthetic media, and we started dabbling with those types of technologies and video tools, and I had my own “digital double” that would speak to the group. I sent a recap email one time with it and recorded the recruitment video announcing the community of practice with that tool. We all dabbled with it and had these really hard and powerful conversations.

One of the things that resonated most with me was how much people felt in their gut that something was wrong. I mean, that sounds slippery, but multiple people felt it.

I’ve thought long and deep about how I would use these technologies. One of the approaches to humanizing is to create a video with your voice narrating while students look at images or slides, with other design elements that make it effective. I think using a voice bot for something like that is something I personally would consider, because there are so many times when only one or two words change in the narrative. Instead of having to sit down, make sure your environment’s quiet, and re-record the whole thing, you could just swap out, change the words, republish it, and the video is complete again — with captions, of course.

So there are opportunities. It’s just that we have to be really sure that students know who we are and that we are there, and be transparent about what we’re using. Explaining to a student: I used this tool to produce a simulation of my voice, and you’re going to hear that along with my real human voice, and they’re serving different purposes. I think that’s really helpful.

Lark Park: There’s so much about this that evokes questions of authenticity — what is important about it, and where the line is when things start to feel like counterfeit or cheating even. Where is that line now, and will that line shift in the future as these tools become more ubiquitous? [Will it become something] that people, again, come to take for granted, as we talked about at the beginning, about how it was amazing to see a message on a computer in real time, and now we don’t even think about it?

Michelle Pacansky-Brock: Yeah.

Lark Park: But this is deeply unsettling at some level, if you dislike change. Maybe it’s exciting for some people to stretch the bounds of what’s possible. What you’re describing actually sounds like a productivity gain. You don’t have to redo the whole thing manually. You can just cut and slice it. And maybe transparency is the thing we have to fall back on, as you said — to be upfront about the use and see what comes back.

Michelle Pacansky-Brock: Right. And going back to where we started — you were calling out a paradox: that I have spent my career focused on using technology to bring people together, but the flip side is that technology has also pulled us apart. We have to live in that space and recognize both sides of it.

I keep thinking about the work of Sherry Turkle, whose work I pushed back on about 10 years ago and have since revisited. I understand her perspective so much more deeply now, probably because I’ve lived in this space for so long. About 10 or 11 years ago, she wrote a book called Reclaiming Conversation. A decade ago, the data was there showing that because of our use of texting and phones, meaningful conversations are going by the wayside.

You brought up the transactional nature that is often the mindset that governs education or teaching. A lot of times we hear about human interactions as being transactional. It’s just an exchange of information. But when we look at the way humans develop, the way we have evolved over millions of years, [it has] been through interactions and conversations with other people that we start to understand how we fit into the world, what’s important to us, and what we value. Over these years, we have been using technology to mediate our conversations, our interactions and our relationships with other humans. But now, more people — particularly young people — are actually relating to and interacting with technology. The technology has replaced the person. And that, to me, is so concerning.

I see this as the next level of humanizing: building in a sense of awareness about the immense data we have on isolation, loneliness, depression, and anxiety — all different things, but intertwined. And how we can practice ways to restore our connection to others, our relationships with those who are important to us, our sense of community — building that in, and finding awe in the world.

I recently read Dacher Keltner’s book on awe, and it’s amazing. There’s research [about] looking up at a starry sky at night and having that feeling that there’s something bigger than me out there. I’m connected to something bigger. I’m sure there are people who would shake their heads and think, how does that fit into my class? But I used to teach art appreciation, and I’d have students wake up before sunrise — or stay out before sunset — and spend an hour and a half watching the way their surroundings changed: how the light changed, how the color changed, what happened to the shadows. I did that because I had a teacher do that with me when I was learning how to take photographs, and it sticks with me. I remember that morning vividly. Those are the things that can make someone excited and feel good and build their overall well-being. I just hope teachers recognize how much power they have to make change.

Lark Park: Can I stay on that — on well-being? One of the narratives [developing again] in higher education is that we should be done with coddling students, that we’ve coddled them too much. It’s a counter-narrative to the last decade, which was: we need to do better to meet students where they are. And now it seems like it’s shifting again to: they’re too coddled, they’re not learning.

A lot of your humanizing work is about the message of care, well-being, and human connection. How do you see the current moment for families, faculty, and students in this conversation?

Michelle Pacansky-Brock: The first thing I would say is: if anyone thinks humanizing is coddling, that’s not what it is. It is always about holding all students to the same high expectations. It’s just about providing the support that many students need to get there. If we, from the beginning, could be invested in supporting all of the students that we actually have instead of hoping for these students that we wished we had….

Jesse Stommel and Sara Goldrick-Rab — years ago they wrote a great essay about that. We’ve got to design our classes and teach for the students that we have. The reality is that the students we serve have been changing over the years. As we become more inclusive and committed to equity and closing the equity gap, we have to reckon with the fact that the way we teach privileges certain students and leaves out others. So it’s up to us to determine how we’re going to think about our class. The way we think about students, the way we think about teaching is where everything starts, and that’s where equity work starts too.

I think it’s really important to start from a place of equity. What does this mean to students and their families? My dad was born in poverty. He was one of 17 kids to the same mom and dad. He was the only son to graduate high school and the only child to go to college, and it’s because he left the East Coast and came to California and enrolled in Porterville Community College in the 1950s, gaining free access to higher education.

Even though I was born into privileges that he didn’t experience until later in life, that’s part of my story. And when I think about the coddling narrative, we’re losing sight of [the fact that] higher education is still the greatest mechanism to change people’s lives, for generational social mobility. We’re not just changing lives — we’re changing future generations. That’s something I’m really passionate about.

Lark Park: Can I ask why you chose faculty professional development — faculty professional learning — as such a focus?

Michelle Pacansky-Brock: I’m not a believer that someone chooses a career and that’s what sticks. I started out with a degree in art. I started at an advertising company that I hated and I quit. I went to Europe (privileged) with a backpack. I remember standing in front of a painting in a museum in Italy and deciding I was going to go back and get my Master’s in Art History. And I did. I was incredibly fortunate to get a fellowship to cover my master’s degree, something I’m always grateful for.

When I was in that program — that assumed students were going to go on for PhDs — I realized I didn’t want a PhD. I realized I loved to teach. I had my first opportunity to stand up in front of students, and while I was scared to death — I think I prepared about 18 hours for that one hour — it was the most exhilarating, amazing experience. And then I had students come to my office hours, and I just loved it so much.

When I started teaching, I really became passionate about online teaching, because I saw how good it could be, and the feedback that students gave me just lit my fire. At that time I wasn’t thinking, “I want to go into faculty development” — my eye was on supporting online teaching. That was the thing I wanted to gravitate toward and find my place in.

Lark Park: I was going to ask you what the most gratifying part of your career has been, but I feel like you just named it — and named the moment, even. Do you want to add anything else?

Michelle Pacansky-Brock: I want to go on record and say that leading the humanizing online STEM grant with my amazing colleagues — Mike Smedshammer, Kim Vincent Layton, Di Xu, and many other people — that was incredible, to be able to collaborate with these amazing humans, and just have space to talk…. We didn’t know what was going to happen when we applied humanizing to STEM. We just went for it. We developed the model and the academy, and it was very successful and so rewarding.

I still go to conferences, and I still have folks who come up to me and say, “That changed my life.” And that warms my heart. So, just having that opportunity, making that impact, knowing that all of those teachers are thinking about online teaching differently, that their mindsets have shifted, they have new skills to bring into their classroom, and that students are benefiting from that – that’s amazing.

Lark Park: Okay, final question. If you had one wish for how higher education could transform over the next few years, what would it be?

Michelle Pacansky-Brock: It’s pretty simple, and it’s the same wish I would have had 10 years ago. I wish every student, every human had access to higher education — and that they’re able to achieve what they want to achieve, that they learn deeply, and that they walk away feeling like they were treated with dignity and respect.

Learning Lab Director Lark Park interviewed Mills to explore what kind of writing still matters and why, the relationship between AI and OER, why she chose to consult for OpenAI (Mills was the only education specialist recruited to test GPT-4 pre-release and report on educational impacts), and what regulation and guardrails we (in the education community) might need as AI enters its next phase of development.

Lark Park: What were you doing prior to the debut of ChatGPT — that was November/December of 2022, right? I’m curious how your life has changed, given how much work you’re doing now that’s about AI. Can you talk about that as a watershed event?

Anna Mills: Oddly enough, I had my watershed moment about six months before ChatGPT launched, when I was playing around with its predecessor, GPT-3, and realized the magnitude of what was happening. I couldn’t drink tea for about a month — no caffeine — because I was so intensely, physically responding to this new reality and realizing how much this was going to change writing, writing instruction, and higher ed.

Before that, I was in the world of open educational resources — excited about creating them, collaborating, building on that work, and talking to people on social media about OER.

Lark Park: Did [AI] take over your life at that point?

Anna Mills: Absolutely. I jumped onto social media that summer and discovered all these amazing people who were already talking about AI in education. Then in August I thought, we need a resource list. I looked around, figured nobody was doing one, so I’d just do it myself. I was right in the middle of creating and promoting that resource list for the Writing Across the Curriculum Clearinghouse when I got drawn into consulting for OpenAI that fall.

Lark Park: How did that come about? How did they find you?

Anna Mills: I had gotten onto their user forum of people using their free playground to explore engaging with these systems before there was a chatbot. Not very many educators were comfortable with that environment. It was simple enough to use, but the interface really seemed built for coders. Still, they had a community forum where people were exchanging ideas about projects involving large language models.

I collaborated with an engineer named , gave him some materials from my OER textbook, and said, let’s see how well these systems do on critical analysis of a text. I was just doing my open-educational-resources-style thing in that space, and I think that’s [when] the head of the red teaming effort at OpenAI, Lama Ahmad, reached out to me.

They didn’t have a systematic effort to look at educational impacts, but they did want to include it in a larger assessment of risks and possible downsides. At first I said no — I don’t want to be compromised by association with the company. But she circled back and said, are you sure? You don’t really have to commit that much. I decided it would be an interesting education, and that I could support their efforts to be realistic about the risks and downsides without necessarily endorsing everything the company does.

Lark Park: So your involvement with them predated the public launch. It took over your life about six months before most people even had access to it. Has it stayed that way? Is it AI all the time, in terms of the work you’re doing?

Anna Mills: I’m still deeply engaged as a writing teacher, thinking about writing pedagogy, and not all of that is about AI. There’s plenty we do that doesn’t involve AI, and there’s more urgency than ever around that work. There’s a deep joy in my teaching that isn’t all about AI. But in the sense of sheer intellectual and moral engagement, most of my effort is around AI because I do think we’re in the middle of a huge shift in our culture and civilization, and we need more people engaged in guiding that shift.

Lark Park: What does that [shift] look like? How is it different than what we’re in right now? What are the biggest things that are going to be different? And [over what] time horizon?

Anna Mills: I don’t think anybody knows exactly, and I don’t think there’s necessarily one transition to one stable state that’s going to come after. That’s an easier way to conceptualize it, but it doesn’t seem very likely to me. Unless we hit some immutable natural law that limits artificial intelligence, I think it’ll continue to evolve.

I’m agnostic about how far it’ll go. The people who know the most about it are completely convinced we’ll see artificial general intelligence within ten years — even the skeptics.

Even with what we have right now, we haven’t really seen the impact throughout society yet. We’re going to see transitions in all areas, looking at what do we want humans to do, what do we want AI to do, and how do we steer that — in all domains including things we think of as care work, because these systems can simulate care and emotional work even if they don’t actually experience emotion. Depending on how fast robotics develops, we’ll face those questions in the physical world too. But what’s immediately foreseeable is that in thought work, these systems can do a lot of what we do, and we’ll be negotiating hybrid forms of control and labor, disruptions in the labor market, big shifts in how power works, and destabilizations in things like cybersecurity. [Given] the threats of AGI coupled with climate change, we’re looking at a lot of potential instability. It’s very concerning, to say the least.

Lark Park: I think what you’re tapping into, in terms of the rate of change, feels new. And anxiety-provoking, because it’s not just one change in one sector. It’s broad and it’s fast — faster than what we’re used to. “Anxious” is top of mind for me. Where would you place your attitudes and beliefs about AI?

Anna Mills: I’m somewhere in the middle. I enjoy thinking about and working with AI — there’s a part of me that connects with it. And I share a lot of big concerns with critics of AI. I do think there are possibilities for really bad outcomes. I also don’t think we should assume that just because there are problems, every outcome will be bad.

There’s a lot of uncertainty, and I think the ethos of open educational resources — putting your work out there, letting others build on it, collaborating, focusing on the ongoing process rather than the immediate end product — is actually really helpful for a time of anxiety and uncertainty.

I’m excited and hoping for major benefits. I also see big downsides and am profoundly worried. And I think that’s actually true of more people…. People assume there’s more polarization than there is. I talk to a lot of faculty and to friends in the AI industry who have very mixed feelings — hopes, fears, concerns, and a sense of overwhelm. As humans, we are able to hold contradictions and uncertainty and good and bad.

Lark Park: I want to go back to open educational resources, because that’s where you spent a lot of your time and career before AI entered the scene, and you still spend time on it now. I’m trying to wrap my head around what each means for the other. Even the launch of ChatGPT in 2022 — making it available to everyone for free — was a tremendously powerful move to engage the broadest possible community of users, which is simpatico with the open movement. Tell me how OER impacts AI, and how AI impacts OER.

Anna Mills: I just gave a keynote for OpenCon Ohio on the power of open in a time of AI. My message is that the ethos of open, and the practices of creating openly licensed resources and adapting them quickly, are really helpful right now. They’re allowing us to respond to AI more collaboratively as educators. And AI is useful in some OER processes, so it can be genuinely supportive.

But the concern about AI slop is there with OER too. There’s a chance that people will ask, why do we need good educational materials if we can get immediate, free answers from large language models that sound so authoritative — just like a textbook? Why do we need human-created pedagogical materials?

I think that’s really dangerous. Students understand that [absence of human-created materials] as a lack of care, as a turning away from them, from the human relationship that’s central to education. An OER textbook is like a letter to students: Here’s what I understand, here’s what comes out of my experience and expertise and time with students, here’s what I want to share with you — what do you make of it? How would you build on it? How would you apply it? There’s a message of care in that, and those messages are a big part of education and what motivates students.

Also, something that just looks like a textbook is not necessarily accurate, reviewed, wise, or framed in a way that’s meaningful and purposeful. Without transparency and guardrails about what’s AI and what isn’t, these educational relationships get undermined. Students shrug and say, “They’re just using AI anyway” — and they take that as a message of instructors not caring, not showing up, not seeing them as students.

[That said]… agentic AI is an incredibly powerful tool that can speed up many aspects of working with an OER textbook — with accessibility, converting between formats, making OER available for multiple modalities per Universal Design for Learning, doing research, identifying where we need to update, working with multiple technical systems. It’s not perfect, but it changes what’s possible in some fundamental ways that are really helpful for OER, because we usually don’t have the funding to do enough. So it’s a tool.

Lark Park: To the extent that there is an OER community [or communities], are there similar strands of thought about OER’s relationship to AI?

Anna Mills: There are definitely rich conversations, and some different approaches are emerging. I have colleagues who’ve done hybrid textbooks that are part AI and part their [own] writing (I’m thinking of here). has come up with the idea of generative textbooks — a collection of prompts that lead students into different educational experiences related to OER materials.

More people, though, are still exploring cautiously. There’s a general sense that we need to be transparent about what’s AI and what isn’t, and that there are potential uses and excitement about exploring that terrain. There’s not as much clarity about what could be lost — that’s the part I want to emphasize. But the whole OER community is abuzz about AI.

Lark Park: Have you heard or felt more resistance to OER because of AI — the concern that the more you share openly, the more AI will simply extract your work and feed it into this machinery of never-ending data-consumption? Is there any backlash, or less willingness to share?

Anna Mills: You would think that would happen because it would be rational to think that whatever you put up is going to become training data. But I haven’t heard much about that. I think the folks who tend to be worried about that were already not necessarily into the idea of putting their work up as OER for adaptation. They were maybe already resisting because they believed in being paid for their intellectual property and were concerned about who would take it and benefit from it.

I think folks who were already interested in OER are just as interested. There was already a divide, and I don’t think AI has changed much who’s on what side of that divide.

But there is backlash against the idea that AI is using everything as training data. ASU put out these materials that were based on avatars of professors without their consent, using their data – and there was definitely backlash. That’s where you see it: my institution now has my data, and what are they going to do with it? Maybe they’re going to think they don’t need me to teach this, since they have all the data from my past online courses. That’s more of a concern than AI training on OER, which was always meant to be shared.

Lark Park: Speaking of backlash, you’ve probably seen the same stories I have about student backlash, speakers getting booed at graduation ceremonies, and so on. Are these just blips, or is something growing in terms of student backlash about AI’s role in education and society more broadly?

Anna Mills: I think there is some growing backlash. We’ve seen it in the [project profile on Learning Lab’s website: Peer & AI Review + Reflection Project] — a few more students are opting out of AI feedback than previously. Still not the majority, but a few more. At my son’s high school student film showcase, one of the lines that got the biggest applause — an intense roar — was the credit on his film that said “Made without the use of generative AI.”

I do think we’ll see populist backlashes against AI, especially since entry-level white-collar jobs are the most threatened. College students, both because of job fears and environmental concerns, have a basic push for agency that I think is very healthy. There’s a sense of “We didn’t ask for this, we didn’t plan this, we don’t have enough say.” Faculty feel that too, in a different way. It’s healthy that people want more of a say. I don’t think it’s realistic that anyone is going to shut down AI. But it’s a healthy impulse toward more democratic oversight, more guardrails, more agency.

Lark Park: In terms of higher education broadly — you’re in California, with enormous systems of higher education. You teach at College of Marin, but you consult with faculty throughout the UC system, the CSU, the community colleges, and in other states [and countries] as well. Do you have a 60,000-foot perspective on where different parts of the system are? Are four-year institutions ahead of two-year colleges, or is California way ahead of other states — or are we all grappling with exactly the same thing?

Anna Mills: I do think California has done a lot and has a lot to be proud of. The SUNY system in New York has also done a lot systematically. I’ve been really impressed by the Idaho statewide AI program run by . Beyond that, I think it’s dependent on who’s at the institution and how things have unfolded there — it’s really variable.

Community colleges have an advantage: we’re already mostly focused on pedagogy, and when the need arises to rethink pedagogy and assessment, that’s already what we’re talking about in department meetings and on sabbaticals. There are no TAs, and classes are generally not huge lectures, so that’s a strength. But we usually don’t have well-funded centers for teaching and learning, which larger institutions have.

Australia has come up with some incredibly insightful, useful, and practical research and methods — the University of Sydney’s two-lane approach, the “Swiss Cheese” model of layering multiple approaches for academic integrity, ’s work with the AI Assessment Scale. Australia is really the place I look to as a model.

Lark Park: Are we missing any big stories related to AI — or really important ones? I often think we obsess about specific things that may not even be the right things to obsess about, while something else goes completely unexamined. Right now the dominant story has been academic integrity, the fear of students outsourcing their critical thinking. Are there stories you think, nobody’s covering this, but it matters?

Anna Mills: I think there are two. One is the idea that AI can help us connect to other people’s ideas — that it can be a form of research assistance, help with fact-checking, help us improve the quality of our own thinking, in ways that are grounded in human sources. It can be used not as an authoritative source on its own but as a thought partner that helps us link with other human ideas. I take this from and his piece in The Atlantic. I don’t think his work gets enough attention. He’s been showing the incredible capacities for things like fact-checking, identifying sources, and helping us see nuances — essentially, the idea that AI can help us slow down and think better if we choose to engage with it in those ways. We need to design systems that nudge us in those directions rather than encouraging us to just shut off our brains. AI doesn’t have to mean “do it for me.” It can be a stimulus to our own sense of agency and our own thinking.

The other thing people are just starting to get their heads around is that AI is now agentic — it uses tools, it works like an assistant, it’s not just a chatbot. And it’s incredibly powerful as an assistant, as well as flawed and risky. We’re still grappling with the implications of the chatbot model, so we’re not yet able to think about how to regulate agentic AI, or how to know when something acting on the internet is a human and when it’s a bot.

Lark Park: I do worry that because we’re not paying enough attention, the development moves well beyond what’s safe or acceptable while we’re focused on the last thing.

Anna Mills: That’s already happened with agentic AI in education. The companies have quietly allowed their agents to go in and complete learning activities on behalf of students. Other educators and I have been testing this and showing it in YouTube clips — you can ask Gemini to take a quiz for you, to fill out a discussion post, to complete assignments in your learning management system. You can say: find out what my homework is and do it. And most of the time, these systems will comply. They will not refuse.

We have run into a really urgent situation where we need some oversight of these agents. That’s something I’ve worked on with the Modern Language Association, on a task force that produced a statement — signed off by the executive board — calling on AI companies, LMS companies, and lawmakers to support educators on this issue, to protect learning spaces online and protect human learning.

Lark Park: Can I ask in terms of responsibility, do the major AI companies just need to control for this? Or is it that institutions just need to respond with more in-person proctored environments? What’s the hoped-for response?

Anna Mills: I don’t think we should count on one entity doing one thing and the problem being solved. We’ll probably have to layer approaches. But it starts with the principle that agents shouldn’t get to pretend to be humans in spaces that are really for humans — like a learning space. The low-hanging fruit is for companies to label their AI agents as AI, so that a learning management system can say: you’re a bot, you don’t take the test.

So that’s a first step, and that’s actually something Amazon has won a court case around, against Perplexity — Amazon argued they don’t want bots shopping for users. If Amazon can say that, I think learning companies should be able to say they don’t want bots completing their learning activities.

And yes, we will need more in-person proctoring in the age of AI, along with other solutions — maybe lockdown browsers as a stopgap, for example. But it needs to be all-hands-on-deck, and not just a focus on what the individual teacher should do to redesign all of their pedagogy.

Lark Park: The Amazon-Perplexity case — you posted about it. Do you see it as an important line in the sand?

Anna Mills: If there’s legal precedent establishing that a website shouldn’t have to welcome a bot if it doesn’t want to, we should certainly apply that to education before anywhere else.

We’re going to need to identify what’s an AI agent and what isn’t, monitor what those agents do, and reduce related security risks. On a broader level, we need regulation of the agentic AI. I think people in education tend to throw up their hands and say we can’t make these companies do anything — but maybe we can. Maybe there’s legal precedent. I take it [the Amazon case] as a sign not to give up on systematic regulation.

Lark Park: I want to talk about writing for a bit. You’ve been a professor of writing for a long time. Do you think writing has been forever changed by large language models? Has writing passed a point of no return?

Anna Mills: We have different choices available to us when we write or when we want to produce text. That landscape is forever changed — yes. We have different options now. But what writing always offered us, it still offers us.

Writing processes are still going to be immensely important and useful if we want to develop our own agency, make our own decisions, think clearly, and use words to be in relationship with other people. We’re still going to need to do a lot of our own writing. Sometimes that might look like prompting; sometimes we might be using AI for various forms of assistance or be in dialogue with AI. But we’ll also need protected educational experiences where we do our own writing, using other strategies, because writing processes are processes that help us think and that help us make better decisions and connect with one another.

[Writing skills] are actually more important than ever, because you don’t need to know Python to build an app with AI. But you do need to clearly describe what you want, and you do need to read the results and critique them. Writing for thinking is more relevant than ever. Writing is a place you can go for renewal, for inquiry, for the biggest, deepest questions — not just about who you are and what you’re here for, but practical ones too, like what’s the next step in my business? A writing process might help you think that through.

So yes, our options have changed. But what writing offers is still there, and we still need it.

Lark Park: I think it’s fascinating that you distinguished [the process of] writing from the production of text. Were they always different, or have they diverged only recently?

Anna Mills: AI brings to light that there are many areas where we don’t really believe in the value of the text we’re expected to produce. Think about workplace writing — reports and other things people are tempted to outsource to AI because they don’t really think they’re getting something out of the process of writing them, and they’re not that invested in communicating something to someone. It’s pro forma.

But if you’re using AI to create the report and the reader doesn’t know if any human stands behind it — at that point it’s lost the value it could have had for readers.

I think there was always a challenge there. You’re asked to produce writing for a purpose — like a program review report. It would be useful to reflect on your program and how it could be improved. But in practice it starts to feel pro forma. How do you return to the initial purpose of that as meaningful communication? AI highlights that tension, and it highlights the need for motivation and purpose.

Lark Park: What you’re raising is this question of when you need a human — or a group of humans — to stand behind something. And occasionally the fear is that students will come to think AI is smarter, or more reliable, than humans. Is that a real concern? Is there a point in time where we cross that threshold where more people look to AI as more trustworthy and stop needing a human to stand behind the product?

Anna Mills: In a lot of contexts — social endeavors, organizations formed around particular purposes — intention does matter, even if AI could produce something that looks more brilliant. It’s important to focus on agency and purpose in our AI literacy efforts, and to give people experiences where they notice that AI sounds smart but isn’t aligned with what they actually think or want.

Increasingly it’ll be harder to give examples of AI being simply wrong or stupid — although not impossible. We’ll have to show where AI says two contradictory things. At some point, you want to live your own life. You have to be making your own judgments and have the mental capacity to do that, even if AI is a thousand times more slick and sophisticated than you are. Because yes — I think it will become more and more tempting to just trust it, because it seems so much smarter.

Lark Park: What is your best advice for faculty and students navigating a world where AI is just part of it — and maybe a big part?

Anna Mills: Openness and engagement — and being ready to hold ideas that are in tension. I think you can’t really understand what’s happening fully unless you are interacting with these systems, what that can look like and feel like.

And it’s critical to watch ourselves very honestly — noticing how working with these systems affects us, where and how we want to shape and limit that. It’s like with our phones: there are downsides to living with them, areas we’re concerned about… and [we have to] choose how much to engage and when.

Try to keep your own sense of agency and stay engaged with AI in some way, in order to understand it better.

And try to help shape it because what’s needed is for us not to polarize into people who hate AI and people who love it, but to be together, working to make it less bad and more good.

Lark Park: I love that earlier point [you made] about [the open community’s] collaborative response — how can we as humans influence each other in this, even if our responses aren’t identical.

📜Don’t miss the conversation on Learning Society’s new research brief, “Who Will Train the Already Educated?”

Tuesday, June 2nd, noon to 1 p.m. PT, Zoom, Register here.

US higher education was built to launch careers, not sustain them. But as working lives extend into the seventies and skill needs shift faster than ever, the absence of any serious collective investment in upskilling is becoming impossible to ignore. In this brief, UC San Diego’s John Skrentny and Mary Walshok argue that the solution already exists inside our research universities — in the units known variously as extension, continuing education, professional studies, or engineering professional programs. The authors give them a single name: career sustainers.
Paul Fain (Work Shift) will moderate a discussion featuring the co-authors, as well as Jason Owen-Smith (University of Michigan) and Kathleen deLaski (Education Design Lab).

Read Learning Lab’s interview with Learning Society’s Mitchell Stevens on Substack.

📜Institutionalized Education as Cognitive Offloading

Stefan Bauschard, Substack, May 19, 2026

“It tells them, by the architecture of the school day itself, what is worth their attention: forty-five minutes of math, forty-five of English, forty-five of science, a thinner slice of art, almost nothing of philosophy, civics reduced to a test prep unit, ethics absent entirely, the questions that organize a human life relegated to “electives” or to whatever the kid finds outside the building. The schedule is a curriculum. The curriculum is a statement about what is real and what is decoration.”

…

“It tells them what to want. The reward structure of the school — grades, honors, advanced tracks, GPA, class rank, recognition, college admission — substitutes an external scaffolding of desire for the internal work of figuring out what one actually cares about. Twelve years inside that scaffolding is enough to permanently confuse the two. Many adults never recover. The high-achieving student who arrives at college unable to answer the question what do you want to study without first asking what looks best on the transcript is not a failure of the system. The system did its job.”

📜Button‑pushing explorers: How to grasp that AI agents can do amazing things while knowing nothing

Ji Y. Son and Alice Xu, The Conversation, May 12, 2026

“There is one game that famously stumped these early neural networks: Montezuma’s Revenge. To make progress, a player must carry out a long sequence of actions – climbing ladders, avoiding obstacles, retrieving keys – before receiving any reward at all. Unlike simpler games, most actions offer very little immediate feedback. The game required something like goal-directed, long-term planning.

Early neural networks would try a few actions, receive no reward and fail to make further progress through Montezuma’s underground pyramid. From the system’s perspective, all actions looked equally useless. But researchers made a breakthrough by changing the feedback signal. Instead of rewarding only success, they also rewarded the system for doing something new. The rewards were for visiting parts of the game it had not seen before or trying actions it had not previously taken. This tweak encouraged exploration.”

📜‘The cost of compute is far beyond the costs of the employees’: Nvidia executive says right now AI is more expensive than paying human workers

Sasha Rogelberg, Fortune, April 28, 2026

“…Analyzing the technical requirements of AI models needed to perform jobs at a human level, researchers found that AI automation would be economically viable in only 23% of roles where vision is a primary part of the work. In the remaining 77% of the time, it was cheaper for humans to continue their work.”

“While AI may cost more than human labor today, there will be warning signs of a tipping point toward AI’s economic viability.”

📜There Are Three Ways to Learn With AI. Most People Use None of Them.

Dr. Sam Illingworth, Substack, Mar 25, 2026

“If you manage a team, ask how they are using AI tools. Not whether. How. The difference between delegation and inquiry is the difference between deskilling and development.”

“The generation learning to work with AI will also be asked to supervise it. Whether they can depends entirely on whether they built the skill or borrowed it.”

Guest:

• April Lawson: co-founder, Insight Debate & Dialogue; former director, Braver Angels College Debate and Discourse Program

Also available on: Apple / Spotify

CHAPTER 1 [0:00-6:37]

Taiyo: Welcome back to My Robot Reacher. I’m Taiyo Inoue.

Sarah: And I’m Sarah Senk. And what a time to be back just a week after the CSU announced a deal to provide ChatGPT-EDU accounts to students, faculty, and staff for another three years.

Taiyo: Hooray!

Sarah: You know, I thought Taiyo, you would be a little more excited than just like a lukewarm hooray.

Taiyo: That was lukewarm? I mean, I don’t know. I thought that was pretty warm. No? Listen, if you want some strong, strong emotion, I mean, I haven’t checked myself, but pretty sure, just go to social media. You’ll find plenty of that.

Sarah: You know what? I don’t think I should go to the place where you presented a perfectly reasonable explanation of how using AI in your course design and university service work has improved your life. And someone told you to kill yourself.

Taiyo: “Oh my God, I forgot all about that. Oh, yeah….

Sarah: Why are you laughing [laughs]

Taiyo: Well because it’s hilarious! Wait…okay let me tell the story. Someone was was inquiring just generally on Twitter or X, however you want to call it, what’s a useful application of AI? And I replied with my anecdote about using agentic coding tools to automate the upkeep of my Canvas course shell. And, you know, it was a very innocent tweet. I put it out there. And I got a reply saying, your life sounds really mundane and boring. To which I replied, well, that was kind of mean. I hope that the hate in your heart dissipates, to which they subsequently replied, if I wanted to be mean, I would tell you to kill yourself.

Sarah: I can’t believe you forgot about this. I mean clearly, it’s burned into your brain.

Taiyo: I mean, listen, when you remind me, yes, it all comes flooding back. But like I should have I should have this example queued up as the kind of toxic discourse which social media supports and provides.

Sarah: Right, right. Yeah true But I think that what also also feels so exemplary for me about social media interactions because it starts with somebody asking what seems like a good faith question: It was along the lines of “help me understand - those of you who think this is going to have positive effects in higher education, If you’re a teacher or professor, like, what am I missing? What is what is a way you’ve actually used this to improve your life?”

Taiyo: Right

Sarah: But the thing is, it’s not actually a search for understanding. And that kind of discourse deeper entrenches people into a kind of stereotypical binary of, like, pro-AI or anti-AI.

Taiyo: Yeah, true. But I think that what this sort of raises is the idea that we really do need real spaces for real substantive dialogue. And, you know, we are all in this time right now navigating this giant social and educational, I guess you could call it experiment or intervention, right? And so we need places where people can actually respond without vilifying one another and caricaturing one another. Like I can see where you’re coming from with that point of view. I can see how your lived experience informs this opinion that you have. And I can see that you are a good faith actor and I can interpret your opinions in that spirit of good faith with charitability. And from that, I can actually learn. We can find some common ground and maybe form a productive collaboration.

Sarah: And also, you know. explaining clearly like, well, this is why I can see that point of view, but also here’s why I am still more concerned about the harm than the benefit that you’re illustrating and vice versa. So that brings us to our guest today, April Lawson, who is co-founder of an organization called Insight Debate and Dialogue. April works with a number of different colleges and universities to facilitate dialogue across difference. And we started working with her a few months ago to create spaces for people across our regional university systems. So like faculty across disciplines and students and staff with a diverse array of opinions about AI to engage in these discussions. And it’s really incredible to be in rooms where the goal isn’t to prove someone else wrong, but rather to gain insight into how others. And I think it also becomes a really good cross-disciplinary gut check, because if you’re, let’s say, a first-year composition professor lamenting the endless train of nonsensical metaphors and em dashes you’ve been getting for like three years, it’s easy to assume that all this AI is good for, you know, is slop and forget that somewhere out there right nearby, AI is solving problems that no human has.

Taiyo: Oh, my God. I mean. So true. I mean, just very, very recently, in fact, OpenAI announced that their models had solved a decades-long mathematical conjecture called the planar unit distance problem. This is one of the famous Erdős problems that captivated a lot of mathematicians’ imagination. And a non-trivial amount of effort was put into trying to solve it. But then along comes… Some version of GPT, and it was able to come up with a novel construction from algebraic number theory to solve this conjecture. And this is a result which, you know, even you have Fields Medalist saying that this is something that’s worthy of publishing in the top tier mathematical journals.

Sarah: I mean, I think the point, things are moving so rapidly and if you… I think you’ve said this before, Taiyo that if you’ve anchored on the outputs of these AI models from like 2023 you’re laboring under a delusion about what they’re capable of. I also will say that I think these types of discussions these Insight Debates and Dialogues help us all see where our delusions and resistances might be and give us an opportunity to self-reflect on why we hold the opinions that we do, which I think is a very important thing if we’re going to make thoughtful decisions about how to respond to this technology. Now on to the interview.

CHAPTER 2 [6:38-20:33]

Sarah: So when we last left off, we were talking about efforts to use AI to surface the collective intelligence of a large group of people. April, we’ve been working with you for several months now to host insight debates and dialogues on AI and higher education at our university and at others in the region. So we wanted to have you on the podcast to talk a little bit about your very human take on collective intelligence.

April: Wonderful. I love that I get to have a human take. That’s ... This is, this is, um, people don’t usually phrase it that way, but I’m happy to represent the human here. Awesome. Uh, and yeah, no, I’m so happy to be here. I’m just loving our work together, and I can’t wait to see where it goes.

Sarah: Likewise.

Taiyo: So April, tell us a little bit about how Inside Debates work.

April: Yes. So when most people hear the word “debate,” they think of either a competitive debate and, like, the kind you do in high school or college, where you’re trying to make make the perfect case with all the right statistics, and don’t let your opponent make any good points, and you just achieve victory. You win. And this is different from that. Or they think of presidential debates or political situations where people talk past each other and are not honest. And this is also not like that. Inside Debates are defined by their spirit, which is of a collective search for truth. And so what I want you to picture is a room of, I don’t know, between 30 and 300 people sitting in, ideally in, like, concentric circles or in - at least in a flat way so that there are no stages. There’s nobody who’s like, “These are the people who really know what they’re talking about, and this is everybody else.” It’s a room of people who are, uh, essentially having a structured conversation and what I would call a constructive conflict about a difficult question. The number one rule I always say is that we ask people say what they actually believe, and so that means we invite their doubts and their, the nuances, right? And the things where they’re like, “You know, I am pro-life, but I do think this one point on the pro-choice side is really good, and I don’t know what to do with it.” And so the sincerity of that is really important. Another thing to know is that everyone in the room is empowered to speak. So all 30 or 300 of those people, including, by the way, you know, the people who don’t think they’re supposed to be speaking. So we’ve had our, our camera guy speak. We’ve had somebody who’s like, they’re actually not a student yet at the school. They’re just checking it out. And I try to emphasize as we go through the debate that, like, actually, we want all the voices in the room literally, and that’s because in the same way that our democratic republic needs everybody to make a good decision and to think clearly about something, we need all the voices to talk about and get somewhere good with the tough questions that we talk about. So yeah, that’s the basic idea. It’s a conversation, uh, that is structured and designed to give people a chance to think together about something hard.

Sarah: We’re gonna talk a bit about AI conversations, but because you mentioned a political example, I wanted to give you a chance to talk a little bit about the origin of Insight Debate & Dialogue, and how it came out of your work trying to bridge partisan divides in American politics, right?

April: Absolutely. I’m gonna start the story a little bit even before that, which is, “I was born….” No. I, um but I grew up in, in a conservative Christian area of Kansas, but my parents, my family inside my house was, like, liberal atheist, like, sort of more on that side of things. And so I feel like I grew up politically bilingual in that I would walk out the door and need to immediately shift the way that I talked about things, and then go back home and, and that would happen again. And so I learned from an early age that people have really different ways of talking and thinking about things, but they have actually important insights that the other side doesn’t have. And so fast-forward, I did high school competitive debate, but then I went to college and there’s something called the Yale Political Union, which is basically a philosophical debate society that uses parliamentary debate to help people… What it did for me is it helped me figure out who I am, who I am and what I believe and what I wanna stand for in the world, and that was the most important institution in my life as a young person, and it was because we had this, this forum, this opportunity to ask each other tough questions, to try out ideas, to say… You know, it’s funny, Tayo, you and I were talking a second ago about how you’re vegan. I was vegetarian at the time, and I got grilled about that, so to speak, and like- I... Sorry, I can’t help it. Um, I, I’m sorry. No, that’s great. I will be nice about the puns. But the... Yes, I did, though, and, like, eventually it changed my… It changed. I realized that, like, I actually should eat humanely raised and killed meat instead of be fully vegetarian, and that’s just one tiny example, but this... Like I said, this thing made me who I am. Mm. And so... And it formed really strong relationships. The strongest relationships I have in my life basically are friends from that group. Then, I walked out into the world and- I’ve done a lot of work on the cultural divide. Some of that was at The New York Times with David Brooks. That was more of a, like, research dimension of it. After the 2016 election, I got sent to report on groups that were trying to solve the political divide, and I found Braver Angels as it was just getting started, which is the largest grassroots nonprofit trying to bridge the partisan divide. And I was asked to join their board, and then eventually they said, “Hey, we think we need some form of debate to engage this stuff, not just workshops. But we’re looking to build relationship, and debate so often seems to destroy it.” Mm. And so I thought about my own life and how the debates that I experienced in college are the things that built my relationships with my closest friends. And I said, “I bet there’s a way to use this.” And so I spent six years developing a program to do that, and it turned out to be especially popular with college students because college students hate labels, especially now. They don’t wanna walk in and say, “I’m a Republican, I’m a whatever.” But they have lots of ideas, and they are passionate about them. And so creating a space where students and faculty and community members that are designed to do exactly this, take a tough issue, bring everybody into a room, and set up a space where you, you can generate insight. And that’s why they’re called Inside Debates, is because the, the kind of thing that I like to help groups do is come to insights that they would never have been able to come to alone.

Taiyo: That’s great. Yeah, no, Sarah and I have been to a number of these Inside Debates now, and I think one of the things that has struck me about the quality of the conversations is that it really provides a space for faculty in particular. We attended ones where it was mostly faculty, sometimes students, and sometimes administration in the room. And we were talking about AI in higher education. And what we found is that faculty are really hungry for a space to be able to express their many variegated and diverse opinions about what’s going on in AI in higher education. Mm-hmm. Uh, because obviously we, we don’t, we all recognize the profound impact that AI is having. And shocker: people are quite polarized on this issue. You have, uh, folks that are deeply anti-AI, wanna ban it outright, and folks that want to reap some of the benefits of the technology. It really feels as though, in particular, the Inside Debate, which as you say, isn’t about, you know, gotchas. It’s not about, like, dunking on your opponents. It’s about trying to surface these kinds of insights. These have just been a fantastic way of bringing out into the air, bringing, bringing, surfacing the insights that are inside of, of the faculty body already. So I know that you’ve facilitated a number of these debates about AI in higher education in particular. Are there any common themes that, uh, come up in these debates that you could, that you could point to?

April: Very much so. And just to lift up something you just said, I think that there’s this idea that college campuses are places where it’s hard to speak up, and that when people say that in the media coverage, that students are the people that they focus on. But my gosh, it is, I think, even harder for faculty to speak up. Because there’s more at risk, they have more to lose, and the tensions are, if anything, more rigid. And so I think that it is always a pleasure for me to have the opportunity to give faculty especially, a, a space to speak, because it sometimes feels like they’re in a pressure cooker. And you have got to have some place that you can actually just talk about things. And by the way, everybody typically became a faculty member because they have a lot of ideas and they care. They care a lot about education, and about society, and about the students and the future. And so it is genuinely a joy to give them a place to, to unleash that. And so I love, love doing faculty debates, and it does often feel like there’s, like, something pent up that, that gets to to release a little bit.

Sarah: So I think I’ve seen you facilitate these dialogues in five totally different locations at this point. There was the Teagle Cornerstone convening at Vanderbilt last fall where we met, which was like 200 faculty members from 80 different colleges. But then there are the ones that you did at Cal Poly Maritime where we had, I think the first one had a mix of about 80 faculty, students and staff. One of my colleagues brought her class. And then the second one that was smaller, that was mostly faculty. We did the one at Berkeley City College that was a little bit mixed. It was mostly faculty, but there... I remember a student and a counselor speaking. And then SF State more recently, that was one where I think students and faculty really spoke in equal numbers. But in every single one of these, it seemed like you managed to create these kind of magical conditions where people were just willing to risk half formed thoughts in front of one another. And I don’t see that a lot. So I’m wondering if you can, you know, disclose your secret on the air.

April: Yeah, I’m happy to, because I want this to exist everywhere. I would also add one more debate to our experience together, which was in your classroom.

Sarah: Oh, yes.

April: Because that’s another very different kind of space, where there were students and, uh, just it was pretty informal, because you let me basically walk in and hijack your class with no warning. But man, the conversation was rich, and that is what happens, is if you- People are often afraid to talk about the hardest things, but we all know that what happens if you don’t talk about them is actually worse, right? Because things simmer, and my colleague, former colleague and, and current delightful friend, Mónica Guzmán, likes to say that the less represented someone is in your community, a type of person, the more fully represented they will be in your imagination - basically they become the boogeyman, right?

Sarah: Yeah.

April: But of course, that’s also, that is conditioned on one thing which is you have to make sure that the conversation goes well. And the number one objection I get when I use the word debate or anything like this is, “Yeah, I had a debate with my uncle a couple weeks ago.” “I can tell you how that went,” right? And so I have spent a long time studying what that is that makes people feel so comfortable, and I learned it experientially. I learned it by being comfortable in this space, and I think that there are a couple things. One is, you know, you guys have seen this, so- I said picture a room, right? Of 30 to 300 people, and there’s ... They’re sitting in concentric circles. What’s also true is that there’s a chair. So I am the chair in the debates that you are talking about, but it can be anybody. And the chair is, uh, moderating. And there’s really only one rule that we use, and it’s from parliamentary procedure, which is Robert’s Rules of Order. You know, they use this in Congress, at I bet a lot of faculty meetings.

Sarah: Mm-hmm. Yep ...

April: Some of the genius is in that system. But there’s only really one rule that we enforce, which is called addressing the chair. And so what that means is that anytime anybody is talking, they’re talking as though they’re talking to the chair. So if Tayo is chairing, um, and I wanna ask you a question about your speech, Sarah, uh, rather than saying, “Sarah, how could you say that about AI... about whatever?” I instead have to say, “Mr. Chair, the prior speaker said that she thinks AI is great. I’d like to know what she thinks about all the people who will be out of work,” or whatever it is. And it looks small, but it’s a huge psychological shift to receive that question because it’s not personal. It’s not an attack. So that’s the first thing is that rule is the key. Earlier when I was working with this format, I used to teach people a bunch of other rules too, how to deal with facts, whatever. But I’ve simplified all that away because that’s really the only one that is super important: You have to, you have to make it a group exploration, not something that is a personal attack. The second thing I would say is that we always pre-plan the first four speakers, and that’s… And not... Well, we always try. Um, but what, what that really means is I always talk to the people who are gonna talk first and say, “Hey, you’re gonna set the tone. You actually have the most important role in this whole debate.” And I do what I can to set the tone right at the beginning in teaching people how to do it. But the people who talk first, I say, “Can you please just help me by admitting something you don’t know, or by mentioning a time you changed your mind, or by telling a personal story or something?” Because that is what disarms people. That is what tells them, “Ah, I’m not talking to someone who is an avatar of the opposition. I’m not talking to somebody who’s gonna hate me. I’m talking to a person who’s thinking and isn’t sure about everything.”

Sarah: I love that. So how many of these debates have you moderated about AI?

CHAPTER 3 [20:33-31:51]

April: Mm. A lot. It’s one of the most requested topics of the last few years. You know, the different years there are different... There’s always a favorite topic, and it’s, it differs. So a few, a few years ago it was immigration. Mm. And then Israel and Palestine. AI, I think, is probably the single most often requested topic, and that’s in a lot of contexts. That’s not just faculty, not just universities even. It’s community members, it’s everybody wants to talk about this.

Sarah: Why do you think that is? I mean, it might be obvious, but I’m curious to hear your take on why people from so many different demographics wanna talk about AI right now.

April: Oh, yeah. I mean, I think that people have an extraordinary amount of hope and anxiety. It’s change, right? And it’s change that’s terrifying because of what it could, the jobs it could take away, the questions it’s forcing about what is a human being. But it’s also, like, there’s so much opportunity, right? And the sky is kind of the limit in terms of, like, what this could do in the world. And so I think those things, and it’s just immediate. It’s in everybody’s daily life. And on university campuses specifically, I think it’s also one of the biggest sources of tension between different groups of people. And so it’s sort of a thing where if you have a chance, if somebody says, “What’s a thing that’s hard to talk about that you wanna talk about?” It comes to mind.

Sarah: Yeah. That checks out.

Taiyo: Yeah. So we’re really interested on this podcast, obviously, about AI and higher education. So in that context, are there any recurring comments or, or patterns that you notice? And in particular, when you have a mixed audience, where you have faculty talking with students, maybe talking with staff or administration, do you see any differences among the different constituencies within higher education?

April: Yeah. There are both micro and macro elements to every conversation about this. So everybody wants to talk about both what does this mean in really practical, immediate terms, and also what does this mean about being a human being? And so those two levels tend to show up everywhere, no matter who you’re talking with. The other thing I would lift up is that there are very much themes, and I feel like you saw this really clearly at the San Francisco State debate you guys were at with me last week. You know, it’s one thing to say, to, like, look at the research that says, “Oh, digital natives, they have, like, a different experience than I’m a millennial, than people who were born in my generation.” But it’s another thing to hear people say, like, “This is just in my world,” right? And so I think in that debate, just to give a little bit of texture there, we had a bunch of different speakers who just the way that they talked about it showed what their world feels like to them in a different, in a way that was different from other people. So for example, I think our first speaker was a young graduate student who was like, “Look, I came to this country to get a good job.” Yeah. “That is the primary thing I’m after, and at the moment, it’s not even... This is super different from what it was when you all were trying to get jobs. I have to apply for 100 things, and I may not even get a response to one of them, and all of them need me to be fluent in AI, really good at it, like, able to move the company forward. And also, by the way, it’s integrated into every little aspect, little crevice and nook and cranny of my life.” Right. And so we had that kind of speech, and then we had, I think, I believe the second speech was from a faculty member who said, “I don’t want students to be using AI any more than they already are. I don’t wanna facilitate this because what’s happening is you’re getting a sameness; people are are no longer learning critical thinking. They are not being forced to actually learn,” basically. I felt like there was a comment at the end about how the one student sort of generalized it as, “We are worried about jobs, and you are worried about learning.” And obviously everybody’s worried about both. And then we had a student, right, get up and say, “I hate AI. Why do I have to talk to AI about everything-” “... when I could just go talk to my professor-” Right ... and get an actual human answer? And then we had people who, I couldn’t read this entirely, but clearly there was some tension around that the Cal State system had an agreement with OpenAI to give access. And clearly different administrators and faculty members had different feelings about that and just different understandings of what the institution is for and what that means about how we should interact with it. So I think that the themes that tend to come up are for young people; I swim in this water. You have got to understand that, right? Second, I’m in a vulnerable position that you may not remember, but try. I have to get a job. I have to stand out. I have to be good enough. Like, this is... That’s scary to me.

Sarah: Yeah.

April: And then among faculty, you get a lot of like, “What is happening to my ability to teach people to think, which is why I’m here? I don’t know how to integrate this into my classroom. I’m afraid that I’m being considered obsolete, but I swear to God, there is some different, something different that I can bring to education than a machine can And then with the faculty and administrators, you see a lot around what’s the future of an institution that is trying to prepare people for the, the next century, and what does that mean about power and about jobs and about service.

Sarah: Right. One of the great things about this format is that the, the conversation opens up to those things, where it becomes clear that the thing that is really bugging one person is-

April: Exactly

Sarah: like, capitalism. Yeah. Right? Or what’s really bugging somebody else is what happened at the K through 12 education system. The thing I love about them is seeing the kind of interconnectivity of all of these different issues. Yeah. And, and then hearing, there’s something again about that lived experience, hearing how it hits, even though, you know, it’s anecdotal, but there’s something very powerful about hearing it from someone who’s saying, “This is how it feels to me in this position right now, in this year- what I’m going through.”

Taiyo: At that debate in particular, the SFSU one, hearing from students talk about their deep anxieties about finding work after they get out of the institution after they graduate, that really resonated with me. Um, not just because of my own past experiences, but I think I’m realizing that the struggles or the anxieties that are being produced by all of the, the deluge of news and, and prophecies about how AI is going to take out the the bottom rung of employment and disallow entry-level positions from you know, climbing the usual ladder of employment and career. That is incredibly stressful for students, and I think it is really important for faculty to hear that because faculty are typically, what, in their 30s or 40s, 50s, 60s, et cetera. They don’t have the same experiences around trying to find jobs that young people of the current generation do. Um, so I think it’s really, really important to understand that. Another thing that you were talking about, um, I just wanna circle back on it ‘cause I thought it was really interesting. You gave a quote from a friend that’s some- something like this, and tell me if I’m getting this right. “The less represented a person is in the community, the more represented they are in the imagination.” Am I getting that right?

April: Yes.

Taiyo: Yeah.

April: I may be butchering her exact words, but that is the idea.

Taiyo: Yes. That’s a beautiful idea because one thing that I think happens, uh, too often, particularly when the discourse happens in places like, like Twitter or now called X, uh- ... which are, have been dubbed the new sort of public square, uh, the new sort of marketplace of ideas, right? We end up making caricatures out of our opposition, right? We end up making cartoonish versions of our ideological opponents, and this does a disservice to the discourse, right, more broadly. One thing that I find really important about these Insight Debates is that they are in person. There’s no Zoom option, right? They’re all in person. You’re all in the same space at the same time, which is, you know, an interesting feature of it, particularly post-pandemic when a lot of things went online. Can you say a little bit about the importance of, just the importance of that fact or that quality of the Insight Debate?

April: So interestingly, although the three of us have not worked on this together, they can be on Zoom.

Taiyo: Oh.

April: Uh, so it doesn’t.. what they are though is synchronous. It’s a right-now-and-I-can-see-your-face conversation. There’s just something about human beings where we can dehumanize each other as long as we do not see each other’s faces, you know?

Taiyo: Hmm.

April: And so if we can see each other’s faces- All of a [00:35:30] sudden, we-- It’s like we forget that people have dignity. You also see that they have complexity, and that there’s something, that they’re living a reality that is just as complicated and confusing and legitimate as yours. And so frankly, you kind of just recognize the foolishness of those simplifications as soon as you hear people talk, tell their own, tell stories, right? That’s partly because every story is more complicated than the, the narratives. Reality is so much more complicated, and human beings are so much more complicated than the caricatures. Caricatures are generally based on the fact that complexity is hard to hold.

Sarah: Yes, of course, and fundamentally, I think we need sense-making mechanisms of all kinds. You know, ones that are, are very low tech, ones that have technology embedded as part of the assignment. We need it all over the place. And if AI is gonna be part of how students learn, I think we also have to be super intentional about investing in, um, so-called human experiences that help them listen and, and disagree, and encourage them to revise their thinking in a community of other people.

April: One of the reasons I have loved working with you two is that your idea about having gen ed be sort of half excellent work with AI, where we, we figure out how do you relate to it in a way that, like, enhances critical thinking, that makes you better, that moves you forward down a line of inquiry faster. But the other half, right, is technology free, and it’s Socratic dialogue, it’s insight debates, it’s something that is, that is only possible if you are a human being interacting with other minds. So to me, what that will do is it will not only make... It will not only enable us to survive the AI transition, it will make education better than it is right now, and better than it was 20 years ago, and better than it was 40 years ago because things have calcified and, and we’re gonna have to get to something that’s purely human in order to keep the core spirit of what we do.

CHAPTER 4 [31:52-40:47]

Taiyo: So when you attend one of these Insight Debates- one of the things that you notice is that out of a room of maybe 50, you’ll have maybe 10 to 20 of those folks who actually speak, right? Who actually get up and, and say, you know, maybe a two to three-minute piece about what they believe about the prompt. What do you think the value is for folks who don’t speak, um, in attending one of these debates? Um, what do they- Mm-hmm ... get out of it?

April: So just to add a little more detail to that, the general structure for Insight Debates is affirmative speech questions, negative speech questions, affirmative speech questions, negative speech questions. And what that ... So it’s maybe one to four minutes of somebody saying, “This is what I think.” And then, uh, another four minutes of people saying, “But what about this?” Or, “I don’t understand this part,” or, “I hear you, but I don’t understand how you can say that given XYZ.” And then you have another person say, “Well, this is what I think.” And then that happens again. And it’s not like it goes Sarah, Taiyo, Sarah, Taiyo, Sarah, Taiyo. It’s like, it goes Sarah, Taiyo, Ali, April, June, Annabelle. Like, it’s, it’s all different people, right? And so, yeah, generally in a 90-minute debate or two hours-ish, you’ll have between 10 and 20 people who get up and say, “This is what I think,” who give, write a speech. But many more people than that ask questions. And then as you say, there are people who say nothing. And one of the things that I noticed really early is that - we do a section afterwards that’s about 15 minutes of reflection where we ask people, “What did you learn over the last 90 minutes, and what did you enjoy?” - and invariably, people who said absolutely nothing will raise their hands quickly to answer those two questions. And what they tend to say, “I really benefited from this. I learned, for example, I didn’t know this side believed this thing for this reason,” or, “I loved hearing this nuance. I didn’t realize we could talk about this issue this way.” Like, they always talk, and so they’re telling me that they get value out of it. And I just think that there is something about knowing that you could speak, that changes how you listen. And that what we’re doing, right? Because what we’re doing is thinking together, you’re part of that. You’re not separate from it. And so, I think that it’s actually quite a rich experience, even if you are not moved at any point to literally lift your voice. One other thing is that a lot of people are not necessarily comfortable at the beginning, but I love watching people... Lots of people get up and say, “I was not gonna speak, but…” And I also love watching people start to simmer in the back, and then, like, you know, you’ll see these people who are sitting in the back, and they’re, like, totally, like, they have their arms crossed, they’re like, “I don’t care.” And then, like, you see them sort of start to lean forward, and then their facial expressions say, like, “Ah, I really wanna say something.” And then eventually they just have to lift their hand up and ask a question. And so people enter at different points of comfort, and there are different levels of engagement that you can use to express that.

Sarah: One of my favorite parts, actually, of the debate that we did at Berkeley City College was there were like 45 people there, maybe from around -we didn’t know most of them - they were from regional community colleges and CSUs and UCs and private colleges. But there were a couple of people there I knew had very different opinions about LLMs than Taiyo does. And I noticed, Taiyo, when you got up to give your two minute take, they sort of clenched up like, oh, God, what’s he going to say? And then when you started talking about how you thought it would be a complete tragedy if students offloaded their cognition completely, and you talked about cognitive sovereignty. They completely loosened up and they started, you know, even tapping to show their affirmation of what you were saying. And so I think that’s also a really great part about being in this crowd of people in real time, observing people, even updating their prior ideas and recognizing that, hey, there are still some fundamental things that even across disagreement you agree on. It brought, I think, a sense of like humility or mutual humanity or something like that.

April: Humility and trust.

Sarah: Mmmm.

April: Because I trust that you’re, you’re thinking about this too. And one of the things that we talk a lot about in depolarization spaces is good faith. It’s assuming good faith, and that’s gone way down in studies that you, you don’t necessarily assume that the other side, the other people are smart or well-intended or-

Taiyo: Yeah ...

April: I don’t know, approaching it with a good, in a good way. And one of the things that happens when you see other people wrestling with something is you’re like, “Oh, they’re trying.” Yeah. “They’re trying as hard as I am.” And- That makes me trust them in a different way.

Taiyo: I think you also, like, begin to understand that you might have differences in how you execute on your values, but, like, oftentimes the values are the same.

April: Oh, totally.

Taiyo: Yes. When we’re talking about, like, folks in higher education, we all care about the wellbeing of our students, and we live to, you know, educate them and to make them stronger individuals. Um, I think we have different ideas about how we should execute on those values, how we… and particularly in the age of AI, when there’s this opportunity slash, you know, uh, this dangerous moment.

April: Crisis? Yes ...

Taiyo: So we just have different ideas around that. But when you realize that your opposition, which you’ve been blindly psychologizing and attributing the most sort of, uh, uncharitable personality, uh, profile of.. when you realize that there, that’s just not the case, you, you realize, uh, again, the complexity of the issue. It really surfaces insights that you probably couldn’t have come up with on your own.

Sarah: That to me too is something we talked about regarding collective intelligence too, and why we find it so compelling, ‘cause it’s this idea about, of, of collective sense-making, which I think is a term I took from you, April, that I use all the time now. But what happens when you surface assumptions and, and generate these, “Oh, I hadn’t thought of it that way before” moments that can actually change the framing and the shape of conversations, and, and I think make people a little bit more humble in their tendency to cling to their beliefs as the only way things should be.

April: In our society, there’s a lot of emphasis right now on individual voices. Like, our whole structure is about individual voices. And even if we call it a conversation, right, like the quote conversation on Twitter, it’s not- In my opinion, that is a conversation that has been stripped of relationship, and relationship is essential. So if you’re not in the room with somebody and you’re... I mean, I suppose you could say that there is a style of relationship on Twitter, but it’s a, it’s a denuded one in my opinion. It’s been gutted of the most important parts of it. And when you can put people back in a context where there’s a relationship between them and there are active social dynamics that they’re responding to in real time, then you get different outcomes. And the simple and sort of banal way to say this is two minds are better than one or whatever. But I also think that part of what’s going on is that people call forth different aspects of one another. In some ways, when I think about how do I want Insight Debates to change people, what I am thinking about is I want to give them an experience of being the person that they want to be. And so there are all these little things that cue people to engage with respect, with humility, with audacity too, right? I wanna hear people say like, “You know, I think maybe Marxism was right after all.” I wanna hear people say that, and, and they, they won’t do that unless they feel it’s safe. And so people love being the person that they get to be in good conversations, right? They love being curious. They love assuming good faith. They love wrestling with stuff. They love this, and it’s a thing that you can make a habit; you can make a habit of being that way. One of the reasons that I love debates that ha- this is, I think, especially true in debates that have multiple kinds of people, so faculty and students, students and administrators, uh, administrators and community members, because there’s just something about those differences that, that I don’t, I don’t know exactly how to say, but it draws out something good in people. And so I think then that the so-called truths or the theories or the ideas or whatever we wanna call them that come out of that context are different, and they’re better.

CHAPTER 5 [40:48-53:12]

Sarah: Going back to something you said earlier, April, about how we’re thinking about integrating AI literacy into general education classes. You know I’ve been thinking a lot about conversations in my field about whether our old modes of assessment and the things we used to help students develop their cognitive skills, whether those are really, like, the, the best possible way to do that or whether there are actually other ways, ways that might even integrate AI. You know, it’s a struggle to teach writing and to assess writing since ChatGPT came on the scene. But I’ve been moving a lot more towards dialogue in my classes and having them record conversations even during class time, like go out and take a walk or wait, take five minutes at the end, record, and then instantly in their CSU provided accounts you get a transcript, and then you can interrogate it, and you can mix and match and have students kind of engage with each other’s ideas that way. And I found that there’s something about... The two things I’m seeing, one, students are coming to college lately with just much, uh, I’d say less finely honed writing skills to begin with. And so writing produces a lot of anxiety. Um, and I’ve noticed students who know the material, they can talk to me about it the morning of the exam, and then they get an exam in front of them, and it’s handwritten, and they can’t write anything because they’re just consumed with the fear that they’re saying something wrong, which is really devastating, I think, right?

April: Yeah. Totally.

Sarah: The, the anxiety around that. But I think up until now, it’s been really difficult to scale oral assignments because you, the instructor, typically have to sit there and listen to, you know, 100 20-minute-presentations, right?

April: Right, right, right, right. Yeah. Yeah.

Sarah: And so I’ve been excited a- about using AI to try and think about how I get them to go out and have dialogues and then preserve a record of it that is gradable by me, right, or, or workable into an assignment.

April: That is so cool.

Sarah: I’m loving it. It’s the first semester I really tried it. I mean, I’ve always done free writes, and I’ve always tried to put in reflection components, and so now I am trying to double down. As I experiment, I try and stick with things that I know are proven. Like having moments for students to reflect on their own thinking after they do an assignment. There’s a lot of literature saying that is a good idea that helps learning.

April: So this is super interesting. I love... This is, yeah. It’s so beautiful to watch the, the ways that people are, again, pulling the human and then out and then making it, like, blend well with AI. That’s a great example of that. I’m curious that, when you say dialogues, is that because you’re catching some of people’s dialogue with themselves? Or how does that…

Sarah: I actually set them out in pairs. and it also solves the problem of if not everybody has a working, you know, has a smartphone.

April: Right.

Sarah: Most do these days, but if someone’s battery’s dead, right, or they don’t have one that has the latest software to do the transcript, anybody with a voice memo transcript. And it’s also great because another good pedagogically thing is getting them walking up outside and taking a five-minute break so that they’re not, like, exhausted in the middle of class. And I think that that, having one other person to confide in and say- “Oh, I didn’t know what she was talking about then either. Let’s ask about that.” It, it’s like, yeah, the way I started it actually was it was them and me in office hours, and then I realized there were things they tell each other that they’re not gonna tell me.

April: Totally. Totally.

Sarah: And I have them all, I keep it anonymous because it’s mostly this is kind of contract grading thing, like you do it, you get the credit for it. It’s all very low stakes things. But so if they have a dialogue with each other- it takes me five minutes to amalgamate all of those and turn them into an outline of here’s a list of top lines of different things that groups were talking about. Here are the things that the majority were talking about. Here are some outlier opinions. And then that becomes the scaffold for the next class discussion. And then people are more willing, especially if they see that, oh, I’m not the only one who thought that thing, then they’re more willing to speak up, right?

April: Yes, absolutely.

Sarah: The, the pluralistic ignorance component, if they can see their thoughts reflected back at them elsewhere, that can, that can be super, super helpful. But the reason I brought that up, though, is something you said earlier about the face-to-face elements of it, because what I’m doing is partially face-to-face, or it’s two people face to face talking to each other. But there’s something about a connection I’m seeing between this and then the kind of othering of, of your typical social media discourse. Right. Or even your typical faculty email listserv. Oh, absolutely. I’ve never seen people be so mean to each other- as on faculty listservs. so mean, and surely so inefficient for working out what needs to be worked out.

April: Mm-hmm. Mm-hmm.

Sarah: But then you have, like, loads of people who are lurkers watching this train wreck unfold in front of you. Nobody says anything. Nobody does anything. This just seems to me like one of those things in higher ed that we should be talking about more.

April: Oh my gosh, totally. Well, and, and what this is, is, um, it’s conflict. It’s, it’s the fact that we’re not managing conflict well, and people are so shy about conflict in our society because they’ve experienced a lot of destructive conflict, and they have very few models of constructive conflict, right? But if you think about it, and this is one of my things I’m obsessed with, so stop me if I talk about this too long. Yeah. But like, I basically think that conflict, if you understand it broadly, it is the thing that creates growth, right? So, um, starting when we are little kids and we have to, we try to assert, “Well, this is what I think,” and then we have to navigate, like, “Okay, but Mom thinks something different,” right? There’s this phrase from psychological research about the fusion of horizons, right? The idea is, so I live by the beach. I can see this part of the ocean to that part of the ocean, but I can’t see the mountains, right? And if you can see the mountains, then our, our talking together means that we can both see the ocean and the mountains, is the, the metaphor. But we probably start with me saying, “The world looks like ocean.” “No, the world looks like mountains.” And until, like, we can navigate that conflict ... So if you do that well, your world expands, right? What I believe in is- teaching people how you move through conflict, right? Like how conflict can be one of the best things in life. And I, you know, part of the reason I loved the debate group I was in is because I got to watch people; I got to befriend people who like have completely different ideas from me. And this is such a cliche, we all talk about this, but like, but I think that you can love someone for the ways that they’re different from you, not just find the ways that you’re the same. Anyway, the next time you see a faculty listserv, let’s have an Inside Debate. Because I guarantee you, there’s good stuff there, right? People care about stuff for a reason. Right. You just have to give them a way to, to engage it that is effective and, and enables us to dignify each other rather than degrade each other.

Sarah: Right. I wanna go back to thinking about AI, moderating dialogues about AI. One of the most beautiful moments I experienced during the Berkeley City College debate we did was having somebody, one of my colleagues in the CSU system, who is very personally against asynchronous online education. And I think right now especially, there’s a lot of reasons to be, because people are having agents do the courses for them, and very hard to verify. But she made a comment from her personal experience that was something along the lines of how asynchronous online education is just a cash grab for universities anyway. And I was sitting next to somebody who, uh, works in the administration at the CSU Chancellor’s office, and who just gasped, right? Um, and she, you know, it looked like somebody, like, just kicked her in the chest or something. And I think if she was... I can’t even imagine what was going through her head, as somebody who, who thinks a lot about this, and is really interested in questions of access and making sure that people who are working full-time jobs have opportunities to get a BA. And before she could speak, somebody raised their hand in the back, and it was somebody who said she was a student enrolled in a teaching credentialing program at the university, and that it, the only thing available to her was asynchronous online, and that this has been a leg up that’s allowing her to fulfill this dream, right? And so it was just one of these things of hearing, again, like anecdotal evidence. Maybe 98% of people in this course are using AI agents and using it to scam the CSU or the c- community colleges for tuition money, which d- was a scam that happened. Yeah. Um, but for all of these, like, terrible stories you hear.

Taiyo: [laughs]

Sarah: Taiyo is laughing about this. For all these terrible stories you hear, I think to, to have the experience of hearing from somebody for whom it’s working well, that to me is more like, let’s try to figure out how to make it more like that. Like, shutting, shutting down the scammers does, should not mean denying educational opportunity to people who want it.

April: You know, I, I agree with you, for sure. But I also wanna take a second and, and back up the, the woman who said that she’s opposed to asynchronous online education. Mm-hmm. Because I think she did something quite brave. That was a provocative thing to say, right? Because we all agree that educational opportunity is so important. But I think that both of them have a real, a piece of the actual truth, right? I think it’s both the case, and I’m just making this up, but I, I’m guessing you guys know more about this than I do, that educational opportunity is essential, and we probably need asynchronous stuff to do that, and that it diminishes the quality of education to some degree, right? And so, like, there’s ... You’ve gotta have rooms where people can say those things. That woman who said that, she said it with passion, right? She said it because she loves education and learning and wants to give students something that’s worth having. And so I think it’s good, even though it will shock people a little bit.

Taiyo: Mm. I think, I think that’s a great example of the audacity that you were talking about, right? Like having people be audacious, say something really potentially controversial like that, and say it with their whole chest, yeah? Just like, completely put themselves out there. And to be in a space like that, because I don’t think anybody really held it against her. It was sincere. And like you- Mm-hmm ... and you know she has good reasons behind that. It’s not - she’s not trying to be mean. She’s not doing it in bad faith, as you say. Uh, no. Mm. It was great. It was a great comment, yeah.

Sarah: So one of the questions I had for you, April, I mean, you talk to a lot of humans. Do you think there is a kind of knowledge that only emerges between people, as opposed to, say, you know, humans and LLMs? And a, a second question related to that is do you think there’s something irreplaceable about human beings thinking out loud together?

April: Hmm. Yes and yes. Uh, and- I, I have to say, the way that questions like this are often answered is in a sort of philosophical way. It’s an attempt to, like, pin down exactly what it is that is human but not machine. And I, I think that this question about what is human and h- what is the difference from what emerges from human relationship versus, uh, human and LLM, or what is the difference between human intelligence and AI intelligence, I think that’s at the level of, of mystery. I think there’s something powerful there that is never gonna be captured perfectly in adjectives or syllogisms or- ... whatever. And so I, I would just say that the best I can do for you in terms of what is it, is that relationships generate things that individuals can never create on their own, and that there is something that we can see. In the same way that, like, what’s that phrase? Art distorts reality into truth. I feel like we can see art and we can see beauty, and can I explain it to you? Probably not. But, like, I think there’s very much something there. I, you know, at the moment, mostly ‘cause I have been talking to you guys too much, I’m kind of optimistic. I kind of think that, like, society’s gonna, is gonna... This is gonna be a painful change, ‘cause change is always painful. But I think on the other side, we could have really beautiful expressions of the human in addition to really beautiful expressions of the human plus, plus the machine.

CHAPTER 6 [53:13-57:52]

Sarah: I love that ending. Beautiful expressions of the human plus the machine. I feel like that’s my teaching ethos these days.

Taiyo: Yeah, absolutely. And I really can’t say enough about how amazing these Insight Debates that we, that we held over the last semester have been. Yeah. And I hope that we do more of them.

Sarah: I mean, we’re planning on it.

Taiyo: Aren’t we? Right? Uh, and you’ll hear all about it, and hopefully you all in the audience can come and, uh, join us for one of these. They are just such fantastic ways of understanding the opinions and the where, where people who are n- who don’t share your beliefs are coming from. And that’s so, so important for humanizing the other side, not caricaturing the other side so that you’re shadowboxing with a cartoon version of your opposition. Let’s bring a little bit of good faith to this, uh, discourse, right? We are allowed to disagree with one another, uh, but we should always be approaching that disagreement in good faith, charitably, um, and not simply a straw man version of your opposition.

Sarah: Yeah, totally. One thing that framing misses is that, like, these are not just, it’s not just about better understanding, you know, what your colleagues are thinking for the sake of that. I think this is also a really important educational experience for everybody who participates- Right ... because this is so new, right? Like, all of us are still learning. Right. I mean, you could say that about everything. Right. But particularly with this, right? There are not a lot of, I would say, I don’t wanna say there are not a lot of experts in this, but the impacts are, you know, so unevenly perceived and, and are in the process of being documented, and I think that it’s really easy to get stuck in the framework of, like, what is familiar to you. And the thing that really struck me about both, it was a comment, it was actually two different people at two different debates said a version of the same comment to me at the end, which was that I didn’t realize how much I needed a space like this, like, just to hear from other people. And I’m not thinking about this, Ty, in, like, an emotional sense, like a group therapy sense, because I know you’re gonna roll your eyes at that. But rather an opportunity, like when people say they, they want space to learn more about AI. I don’t think that means another workshop on, like, how to make a custom chatbot, right? Really what it means is to learn how other people, and not just how other people are using it, but how other faculty are thinking about it.

Taiyo: Agreed, yeah.

Sarah: Right? That’s the kind of thing that I think can open up the possibilities of reimagining what you’re doing in your courses in, in a way that is, like, commensurate for, you know, this time and place. Because as I’m always telling colleagues, learning more about AI does not mean that you have to have your students, you have to even use it in your class.

Sarah: It just might mean thinking about ways to make your assignments more resilient, ways to, like, it’s an opportunity and an urging to think more critically about what we’ve always done in the classrooms and how we can make those things more effective in the conditions that we’re currently teaching in.

Taiyo: Yeah. I mean, the world is changing really, really rapidly. The competency or things that might’ve been really important competencies for being a working professional in the ‘90s might be quite different when we’re talking about the 2020s, right? And so you may not necessarily have to put AI, uh, scaffolded assignments in your curriculum, but you can still be thinking about these kinds of issues. And why would you do that? Why would you put the time and energy into doing that? It’s because you care about your students and their well-being- Ah ... and their future, right? This is why we’re in this business of teaching and learning of our students. It’s because we care.

Sarah: Thanks for listening. My Robot Teacher is hosted by me, Sarah Senk.

Taiyo: And me, Taiyo Inoue.

Sarah: And it’s produced by Edit Audio. Special thanks to the California Education Learning Lab for sponsoring this podcast. If you’re enjoying this podcast, please consider taking a moment to write us a review on Apple Podcasts or YouTube. It really helps new audience members find us.

Taiyo: And if you wanna drop us a comment, please feel free to email us.

Sarah: You can find our contact details at myrobotteacher.ai.

Learning Lab Director Lark Park interviewed Dr. Fox (who is also a two-time Learning Lab awardee with Professor Dan Garcia) after he co-published an (Inside Higher Ed) op-ed, “Blue Books Are Not the Answer to AI” and just prior to him taking students to the Computer History Museum in Mountain View, CA. Below is their conversation about how computer science education and coding have changed, what failure and responsibility in computing’s history look like, what students in CS should be learning and doing today, and why attaining “the sum total of human knowledge” is still what we ought to be encouraging for anyone studying CS.

Lark Park: Thanks so much for taking the time to do this. Human-computer interaction seems to be where a lot of people’s thoughts are. What’s the state of play in human-computer interaction?

Armando Fox: When I say that my research is at the intersection of those fields, that’s a nice way of saying I’m probably not really an expert in any of them. I’m sort of good enough in all of them to find interesting ways to combine them.

What fascinates me about HCI — human-computer interaction — is that it’s the rigorous study of how people engage with technology, and at its best it can tell us things about how technology can be brought to bear, broadly speaking, to make people’s lives better. That could be by automating or simplifying tasks that people don’t enjoy, or by helping people achieve their goals — to learn something or to get something done.

Unfortunately, since roughly the Web 2.0 era, a lot of valuable HCI research has been subverted in the service of what is widely called the attention economy — or the surveillance economy, if you want to be a little more sinister. I remember a colleague from Berkeley — this was in the early days of Facebook — a brilliant engineer who, in an after-dinner conversation, said it was distressing how much brilliant talent in engineering and HCI was being wasted on getting people to look at more ads.

The state of HCI is as vigorous as it’s ever been, and at its best we could be using it to deeply study the engagement between people and technology, including AI. HCI has been around for a long time, and there was a lot of work that really cared about that.

If I think back to the early days of Apple — there was a time when people at that company genuinely thought they could make the world a better place by providing well-designed, tasteful products that actually helped people do things, that were delightful to use, that made life better. A lot of what you hear about technology today has gotten very far away from that. The engagement with technology no longer really focuses on the benefit to the individual. It focuses on whether people will consume more ads, buy more things, spend more time on a site, become more enraged, or upvote something. It’s disappointing that so much attention is being paid to a field that historically has been able to do so much good.

Lark Park: Who’s responsible for this — I’ll just call it overemphasis? And who’s responsible for trying to get it back on track?

Armando Fox: There’s an amplification effect that is different from what it was just a few decades ago. The speed with which a piece of technology can be taken up by literally billions of people is unprecedented. That’s really only been the case since the early 2000s. If you unleash something, it can get much bigger than you expected, much faster than you expected. I think it’s more important than ever — for exactly that reason — to ask: when I make a piece of technology and put it out there and people start to take it up, who is being invited to participate? Who benefits if it’s taken up? Who am I excluding? What are the biases — implicit or explicit — that are built into its design assumptions, and how are those going to manifest when people who are nothing like me start using that technology?

One thing the history of technology teaches us — not just in CS, but really all technology — is that successful technologies ultimately get used in ways their creators did not anticipate. But that’s not a pass to say, well, I’m going to invent something, it’s going to be used in ways I couldn’t anticipate, so I’m off the hook. On the contrary, that means there’s that much more pressure on you to think deeply about the ways you at least could anticipate. I think there has been a general failure — mostly across industry — to do that. And when I say a general failure, I don’t just mean that potentially negative consequences have not been sufficiently investigated. I think there are times when they’ve been investigated, ignored, and set aside with the attitude of, well, that’s not great, but we can probably live with it. I will let readers decide which companies my observations might apply to.

Lark Park: Let’s get a little more historical perspective on computing. Living in California, with UC Berkeley close to Silicon Valley, the lore of the computing industry — Moore’s Law, and so on — looms large in terms of California identity. Is there something about computing history that people ought to know or would be surprised by?

Armando Fox: There’s tons. As a field, we don’t do nearly as much to understand our history as, for example, physicists do. Physicists know the history of their field; they understand when things were discovered, what mistakes were made, and what pitfalls to watch out for.

In CS, it’s fairly common for a new shiny thing — typically a language or a framework — to come along, generating a lot of noise in the blogosphere, and then one of the grizzled veterans will say, well, this idea kind of dates from the 1970s or 80s. We tried it then, there were some reasons it didn’t work, and some of those reasons are now different, so it’s worth trying again — but understand that this idea is not new. From a purely technical perspective, the canonical example is that the ideas behind deep learning have been around since basically the early 1970s, or earlier. We just didn’t have the data and compute platforms to really push on those ideas and see how viable they were.

But more importantly, there’s something that is often overlooked: there was a certain kind of person who was in a better position to make those advances — whether because of privilege, social and economic class, or other factors. A lot of the received history of computing is essentially the story of middle-aged white guys. I think it’s important to understand that the systems that were built, the way they were commercialized and deployed, and the design assumptions that went into them came from one particular perspective. To the extent that computer science ought to be helping solve real people’s problems, the history of the field does not necessarily come from the broadest perspective of what those problems are or what the needs of those people might be.

And then there’s a third thing we can learn from computing history: it’s also the history of business decisions. The only computing artifacts we use are the ones that at some point became commercially successful enough that it was worth someone’s while to produce them. A lot of the boom-and-bust cycles, the ways technologies get misused, the way technologies become the new bright shiny thing and get shoved into places they don’t belong and nobody really asked for — we can understand some of those things from computing history as well.

When I take students to the Computer History Museum [in a couple of days], I want them to come away with the stories at the intersection of computers and human beings — the people who made them, what they were trying to do, what background they came from, and what was the effect of what they built. I want students to understand that context and then ask: which of these lessons could somehow be applied to what we’re seeing today? In the case of students, they’re about to graduate and hopefully land a position they enjoy, or maybe go to grad school — but they’re going to start spending thousands of hours a year of their talent and time doing something in computer science. What’s the effect of that going to be? Who is going to benefit from what you do? Who are you excluding? Is there anything in the history of the field that will inform how you approach this job? When is it okay to say no? How will you know if you’re being asked to do something that goes against your values? People who work in the social sciences think about this stuff all the time. Computer science students ought to think about it more, and that context from computing history is what I hope they — we — can learn from it.

Lark Park: I imagine that kind of perspective does a lot to build students’ sense of agency — that there are real decision points here, there’s an ethics to all of this. People sort of understand that about the present day, and about how so much of this is driven by commercial interests. I’m probably one of those people who, thinking about the history of computing and how things came into being, would have thought about it [the history] more like the invention or discovery of fire — it just was, and could it have been any other way? And you’re telling me: yes, it could have been otherwise.

When we talked a couple of years ago — about ChatGPT, about the advances in coding, how coding had changed and was going to keep changing with ChatGPT, you made this analogy to reading [and playing] music.

Armando Fox: With respect to performing a piece of music — the way we started out doing programming decades ago, say in the 40s and 50s, was analogous to see[ing] a note on a page, figure[ing] out which line of the staff it’s on, which key on the piano it corresponds to, and instruct[ing] the various muscles in your arm to assume the correct angles and press down on the key. As you become more experienced at reading music, that level of abstraction disappears, and you say, oh, that’s an arpeggio — I’ll just play that arpeggio. And if you become an improvising musician, like a jazz musician, there’s an even higher level of abstraction where the notes aren’t even written out. It just says, the style of the song is this, and the harmonies and the chords are these.

Over time, the levels of abstraction for programming have gotten higher and higher. In the early days, you pretty much had to be an electrical engineer, because that was the level at which programming was done. Most people doing programming today don’t know anything about electrical engineering, and that’s fine.

AI-assisted coding adds yet another layer, though it’s a much bigger jump. What’s different about it is that so far, all the layers of abstraction have been formal systems. When you write Python, you have to write it a certain way — it has syntax rules, it has semantics. You’re still formally specifying what it has to do; it’s just that each token or line of code accomplishes much more than it would have fifty years ago. The difference with AI is that you can use natural language — which is imprecise, not formal, and not well-specified — as a way of describing what you want.

The pitfall with that, of course, is that because you’re not using a formal language, it’s possible you’re not specifying things exactly right, or you’ve forgotten something, or there are cases you haven’t accounted for. You’ve described behaviors you want, but you’ve failed to describe behaviors that are bad and should be forbidden.

So the good news is that AI-based coding will enable more people with less training to do certain kinds of things. However, I think it’s a mistake to assume that means people with less training will be able to do everything that trained people are doing today. There are cases where you really need people with deep knowledge doing a very specialized task. But for every one of those people, there are now hundreds or thousands who don’t have that knowledge but can still do something useful.

If I stretch the analogy: there are specialty surgeons, general surgeons, general practitioners, nurses, physician assistants, and healthcare technicians. There are many different ways to add value in healthcare. I think AI is a new and interesting way to add value in the realm of getting machines to do useful things for us. I’m not confident that’s how most people are going to use it, but that is at least what’s possible.

Lark Park: What does that mean for CS students? How is what undergraduates are learning in computer science changing?

Armando Fox: What we’re trying to disentangle — at least in the conversations my colleagues and I are having — is this: for a long time, there has been a body of computer science concepts that are taught in a way that is entangled with learning to code, writing code, debugging code, reading code, reviewing code.

The challenge now is that certain parts of the code production task can be very efficiently automated. So the question is: if we tried to separate the concepts from the fact that those concepts are taught through the medium of writing code, what would those concepts be?

It’s going to be like calculus. Everyone I know who’s an engineer at some point took differential and integral calculus — they did closed-form integrals, they learned all the tricks for doing integration — and then in real life, you never do that. All interesting integration gets done through numerical methods. But your understanding of what integration is and what the limitations of numerical methods are informs the way you use that automation.

How important is code comprehension, and can it be taught without code writing? I think there’s probably more than an incidental connection between that question and the question of learning to write versus learning to read.

The other thing that’s making this difficult is what’s happening in industry. If you look at the companies that are successfully using AI responsibly and improving their productivity, they’re companies that already had good systems in place, including the people skills aspect, which is hugely important. But those systems were put in place and are being maintained by people who are quite experienced.

So there’s a gap: how do you become a senior engineer without going through what was effectively an apprenticeship? That challenge is being mirrored in academia. I just finished teaching a software engineering honors course where students build software pro bono for nonprofits. We encourage students to use AI responsibly, we review their work, we have great discussions about it. But it works because these are advanced students. How did they get to be advanced students? In their generation, they got there by learning to code the hard way.

How do students get to the upper division now, in a world where AI can trivially do all the lower division assignments? What do we need to be teaching them so that even if they write a lot of AI-assisted code, they understand the limitations? What does that kernel of knowledge consist of? That’s what’s changing — we’re trying to identify it and ask how to teach it effectively, and whether there are places where we no longer need code writing as the vehicle.

Lark Park: That’s a big one: if the bottom rung of the career ladder is effectively broken, and industry is going to look for greater experience, how do students actually get that experience? Also, the signals aren’t clear. Some data suggests that software engineering jobs are disappearing. Then another report says software engineers are fine.

Armando Fox: Well, “software engineer” isn’t really a precise job title. If you told me registered nurse jobs were disappearing, that has a very specific meaning, in part because it’s tied to a standardized accreditation exam.

There’s no analogous thing you can say about software engineers. During the boom times of boot camps — learn JavaScript in ten days, we promise to get you a job in six months or your money back — a lot of us grizzled beards were saying: if these boot camps are teaching you superficial skills, your utility is going to be short-lived, because at some point those skills are either no longer going to be necessary, or the tools are going to get better and you won’t be needed for that anymore. And that’s exactly what’s happening. I know people whose title is software engineer, but whose skill set is fairly limited to things that have now been largely automated away.

What part of front-end software engineering has not been automated away? The design part. The human-computer interaction part — the way a human being interacts with the site, the interactions they go through, the way cues are presented visually. That part is still really important. But the mechanics of how you get that to work have been largely automated away. And if you went through one of those boot camps and didn’t come away with fundamental, transferable knowledge that allows you to approach other analytical coding tasks, you’re probably done.

One of my students asked me: if I become a software engineer, is AI going to replace me? I said: no, what’s going to happen is that software engineers who have foundational knowledge and know how to use AI to be more productive are going to replace the ones who don’t.

If you have deep knowledge, if you can think about systems and architecture, if you can talk to a customer and listen to what they say they need and turn that into a technical plan with design alternatives — and if you can figure out how to use AI to make that process better — people want that.

Lark Park: What you’re saying sounds like a real endorsement of a university education. That what you’re going to get is not going to be fast, not going to be shallow — it’s going to be deep and comprehensive. Assuming you’re nodding yes to that, let me ask about anxiety that’s circulating: are students actually learning what we want them to learn? You co-authored a piece in Inside Higher Ed recently called “Blue Books Are Not the Answer to AI.” What’s driven the move to blue books is the concern that students are going to be able to cheat more, and we won’t really be able to tell whether they got the benefit of a true university education.

Armando Fox: Let me start by saying that the people advocating for returning to handwritten exams and blue books are, in my opinion, conflating two distinct things.

One is: we can no longer trust that when students produce work outside of a controlled environment, it’s their own work. That has always been true — there’s always been a gray market of people who will write your reports and do your projects for you. But with AI, the price of doing it has gone to zero. So one concern is whether, in the absence of a monitored environment, you can trust that the work product reflects actual learning.

That’s a perfectly valid concern, and I agree with it. It has always been important to make sure that when a student produces a piece of work, the fact that they produced it somehow implies that they learned the process. AI has sort of blown that up.

However, the second thing is: what is the right way to administer assessments in a controlled environment so you can trust the results? That’s what our piece was arguing for. AI presents an easy way to obtain a credential for now, because the credential is framed in terms of jumping through certain hoops, and there’s now a way to jump through those hoops without doing much actual work or gaining any real understanding. But you’ll still get the credential, and the credential becomes your ticket to economic security.

In CS, that second part of the equation is no longer as true as it was. You can try to AI your way through a Zoom interview if you’re very clever, but at some point you’re going to sit across the table from someone, and you’re not going to be able to do that. You can’t really fake it till you make it in that regard.

When we argued against blue books, we were saying: yes, we need a proctored environment, but there’s a better way to do that than handwritten exams with 800 people in a gymnasium.

I’d also like to frame our argument for computer-based testing not just as a fix for the AI cheating problem. There are real benefits to being able to do frequent, short assessments. It’s a key ingredient of mastery learning, for example, and it’s actually the reason we got interested in computer-based exams in the first place. Retention is better, it’s better for pedagogy, it’s better for student mental health. It’s also better for the quality of life of instructors and TAs.

Lark Park: In full disclosure, your two Learning Lab grants are about mastery learning.

Armando Fox: Very much so.

Lark Park: A lot of things that cause anxiety get thrown into the same stew. The other thing that’s related — but not in a good way [to mastery learning] — is the curved grading issue, which seems to be gaining some steam. The whole grade inflation conversation: everybody’s getting an A, something must be terribly wrong.

Armando Fox: Grade inflation is a problem, but curving doesn’t fix it. Grade inflation without a curve means too many people who got A’s should probably have gotten B’s or C’s. Grade inflation with a curve means not only that, but you can’t even tell what the people who got A’s know how to do. At least without the curve, you could say: the people who got A’s are supposed to know 90% of the material, but there are a bunch of people who know 80% and received an A regardless. That’s grade inflation. Grade curving means whether you get an A depends on what the next person got.

I can’t think of any justification for that — unless the goal of grades is to identify the top X percent of students within a particular cohort. And not all student cohorts are equal, not all instructors are equally effective, and course materials change.

Lark Park: I do think higher ed has been on the train of wanting to identify the top five or ten percent. Part of that, I think, is about self-propagation for graduate school. But I think that tends to have collateral impacts that are less desirable.

Armando Fox: Even if your goal was explicitly to identify the top ten percent of performers, there are many ways to do that. When I talk to people in industry who are doing the hiring, they basically say: grades mostly don’t matter. That’s a little glib, because due to grade inflation, it now looks suspicious if you don’t have really good grades. There’s a kind of expectation — like, what, you didn’t get all A’s? Grades aren’t the signal that people who care are using to identify who’s best.

The signals they’re using are: did you participate in hackathons? Do you have personal projects you do not for credit, not for a class, not for money, but just because you have the passion for building things? Did you try to get involved in research? Are you a leader in an organization — technical or otherwise? That’s how you identify top performers.

Lark Park: So why has mastery learning been such a focus for you?

Armando Fox: Because I don’t think it has to be the way that it is. For a long time, I think the main reason not to do mastery learning was that it was too expensive. We developed an assembly line model of education because we had more and more people wanting an education, and a limited number of people who could teach them. So the kind of individualized attention — lots of practice, flexible deadlines — we couldn’t afford to do that. Our argument is: if you could afford to do it practically and without breaking the bank, why wouldn’t you? Because we know it works. If there’s something you’re not doing that you know would work, you have to ask yourself why. If the answer is you don’t have the resources, that’s unfortunate but understandable. But if technology can help you get closer to doing it — well, that’s an example of a good use of technology. I’m a big proponent of AI in the classroom for exactly that reason. There are things AI can help with that will help students learn faster.

Imagine having an AI conversational assistant or programming partner. In my software engineering courses, we have an AI play the role of a customer so that students develop the skills of interviewing someone non-technical. That’s hugely valuable. You can’t go out and get real industry people to role-play a customer for an hour with every student — that’s unaffordable. But if we can develop that skill in a less expensive way, and we know it’s an important skill, why wouldn’t we work on that?

So to me, mastery learning, once it’s explained in those terms, should be a no-brainer. We know it works; the only question is why we aren’t doing more of it. The implicit answer has been that we couldn’t afford it. We’re trying to wake people up to the fact that now, we sort of can.

Lark Park: I want to go back to something you said earlier about the history of computing — the role of identity, the people who innovated and made business decisions. There’s subjectivity there. I want to ask about your roots, and how they’ve informed the decisions you’ve made — how you engage as a professor.

Armando Fox: Let me give both a little-picture and a big-picture answer. As you mentioned, I do music and theater. Friends and colleagues have asked me: there’s AI-generated music now, AI is writing plays — how do you feel about that? Is it going to write music that’s better than yours?

I never really understood those arguments, because making music and making theater is fun. I can’t explain why it’s fun. I think it’s something deep in our DNA — we like storytelling, we like expressing emotions, we like the idea of connecting with another human being, having emotional experiences, bringing those experiences together. There’s some primal connection in there. So the fact that connecting to other human beings through that medium is enjoyable — whether tech or AI can do that better seems to me completely irrelevant. There’s an intrinsic joy in doing something that connects me with other human beings, and I think that’s fundamental. I try to get my students to see that.

There was a set of videos that came out in the late 90s called The Triumph of the Nerds — not to be confused with Revenge of the Nerds. There’s a segment with Steve Jobs, when Apple was just hitting its early stride…. he was talking about the design of the Apple II and later the Macintosh. At some point as a college student, he had taken a class in calligraphy and developed an appreciation for the joy and beauty of beautiful writing — and because of that, he was so insistent that the Mac would have proportionally spaced fonts. It was the first mass-market computer to really attempt to do that.

He said — and I’m paraphrasing — that good product design is about trying to understand the best things that humans have done in all areas of knowledge and culture, and finding ways to bring those achievements into what you do.

That’s the other thing I try to convey to my students. Music and theater happens to be my thing, but I’m also a fan of reading history, of understanding political economics. There is no piece of knowledge I’ve ever come across where I could confidently say, that’s never going to be useful in my life as a computer scientist. One of my colleagues at Berkeley, Paul Hilfinger — a professor who retired a couple of years ago, legendary in the department — when students would ask what’s going to be on the exam, he would say: it’ll cover the sum total of human knowledge, but with a very strong focus on compilers. At the risk of putting words in his mouth, I think that was his way of saying: yes, you’re here to become a good computer scientist, but you can’t be a good computer scientist without being a good human being first. And to be a good human being, you ought to understand the best and worst things that human beings have done, and ask how you can bring those things to the project you’re working on today.

People ask: do you combine music and technology? Not directly, but music is an aesthetic. Music can be beautiful in different ways — the way Rachmaninoff is beautiful, or the way Mozart is beautiful, where it’s really transparent, or the way Bach is beautiful, where there’s deep structure combined with an aesthetic. All of those concepts have parallels in technology. Code can be beautiful. The way a piece of software interacts with you can be beautiful. But you have to bring the beauty — beauty is something human beings do. Without it, we’re going to build crappy things and become automatons.

Lark Park: Can I just say that everything you’ve said is not the typical view most people would have of a professor of computer science? And it’s exactly the kind of thing students should understand….

[On the topic of things that students should understand:] One of your sites had your “failure” bio…. I wanted to ask, looking at your “failure” resume, but also who you are and all the successes you’ve had — are you happy with your career journey?

Armando Fox: I am happy with the way it’s gone — which is different than saying that at every moment I was happy with what happened next. It’s well known — I didn’t get tenure at Stanford. I’d be lying if I said that didn’t impact me at all. But I came to realize that I have to define success on my own terms.

I talk a lot about imposter syndrome. Students say, oh, I have imposter syndrome — everybody at Berkeley seems better prepared or smarter than me. I say: welcome to my life. We have six Turing Award winners among our faculty and emeritus faculty. Of course I have imposter syndrome — I have it every day. But I understand that I love learning and appreciating the best things people have done, and part of the cost of that is recognizing that I’m not going to do some of those things. On the other hand, there are things I do that bring me incommensurable joy. I can sit down and play the piano and have an experience that no one else can provide.

I feel very fortunate. And I say that having a long failure bio, with more items surely to come — but that’s okay.

Lark Park: As a proponent of public higher education in California, I’m thinking that not getting tenure at Stanford is what brought you to Cal — and that was a huge benefit. You teach so many more students at Cal than you otherwise would have at Stanford, and they [students] are the beneficiaries of that.

Armando Fox: Not just the number of students, though. We do have a wider range of students in terms of backgrounds and socioeconomic conditions. You and I have talked about my own background. My parents escaped from Cuba — it was after the Castro government came to power. It became literally illegal for professionals to leave the country. They left on the pretense of a one-year fellowship in Spain, because they had both finished medical school and been granted the equivalent of a one-year postdoc. On paper, they were supposed to come back after a year. In practice, they knew they were never going to come back. They were fleeing as political refugees, and they were allowed one suitcase per person. They literally walked away from everything — their houses, their friends, everything. Imagine putting some things in one suitcase, walking out of your house, and knowing you’re never going to be in that world again. It’s hard for me to imagine.

What I got from that is: what you know how to do, and what brings you joy — those are things no one can take from you. Focus on those things. That’s the other thing I try to tell my students: whatever background you come from, if you’re here to learn things, what you learn can never be taken from you. My parents, because they had medical knowledge and degrees, were able to rebuild from zero — start completely over in another country where the language spoken wasn’t their first language, with no connections, no social network, really no professional network. And they did pretty well.

So my parents always told me: do whatever you’re going to do, but do it well. Because that’s the one thing that can’t be taken. You might be forced to leave your country, things might be stolen from you, but your knowledge and your skills and the value that you can offer other people by virtue of those — no one can take those away.

Guest:

• Kim Polese: technology executive, founder, CrowdSmart Inc. and Common Good AI; Public Policy Institute of California Statewide Leadership Council member

Also available on: Apple / Spotify

INTRODUCTION [0:00-8:19]

KIM POLESE: There are many other ways to use AI, one of which is to use AI to help humans actually collectively problem solve and learn from each other, and also collectively learn from data, but interacting with each other. And we can use AI to do that at scale.

SARAH: Welcome back to My Robot Teacher. I’m Sarah Senk.

TAIYO: And I’m Taiyo Inoue. And welcome to all our new listeners!

SARAH: Yes, happy to say we had a big bump in subscribers earlier this month, presumably because we got mentioned and quoted in the LA Times!

TAIYO: That’s some nice self-promotion there, Sarah. Well done.

SARAH: You know it!

TAIYO: To be clear, though, the article wasn’t entirely about us.

SARAH: No, not about us at all, but about Cal State Faculty.

TAIYO: Right, true. The CSU, where Sarah and I are professors, released a report on one of the largest studies to date on AI in higher education, and the article was about how faculty opinions about AI are - SHOCKER! - very polarized.

SARAH: Shout out to the San Diego State University researchers behind the survey, whose work - like ours - is supported by the California Education Learning Lab. We hope you too will check out our YouTube channel and leave a review on Apple Podcasts to drive more of that sweet, sweet internet traffic our way.

TAIYO: My god, you’re a publicity monster!

SARAH: Today’s episode is about how to use AI to summon the wisdom of the crowd without the stupidity of the mob. Our guest is Kim Polese, founder of CrowdSmart AI and Common Good AI, whose work explores how AI can be used not just to generate content or outsource drudgery, but to help groups reason together, deliberate more effectively, and work through complex problems by surfacing their collective intelligence. Taiyo, what is collective intelligence as you understand it?

TAIYO: Oh, God, you’re really going to put me on the spot, aren’t you?

SARAH: You’re so good when I put you on the spot.

TAIYO: Okay, okay. I guess I think of collective intelligence as being like a group of minds that come together and function almost like a single distributed mind, capable of generating insights and ideas that no one person in the group could have generated or produced alone. You know, pulling out the intelligence of a collective can be really, really difficult. For instance, maybe... individuals in the group bring different assumptions or different ways of thinking. I mean, just take our collaboration, for example, right? We bring this up a lot, but I’m in math and you’re in comp lit and we have really. different, very different ways of thinking about stuff. I think one of the really beautiful things about our collaboration is that we can kind of bounce ideas off of one another and cast these ideas through our respective disciplinary lenses, we can see beautiful new insights that I don’t think either one of us could have come up with on our own.

SARAH: I love that. So we’re like a hive mind of two.

TAIYO: I mean, “hive mind” kind of has some negative connotations. I’m kind of thinking of the BORG now.

SARAH: I do feel like you have infected my mind in some way. I’ve noticed I use your mathy language for things. If I have an expectation of what somebody’s going to do and they really surprise me, I’ll say, hold on, I’m updating my priors, which I now know from you is something from Bayesian statistics.

TAIYO: Absolutely. That’s amazing. Yeah, exactly. When you come across new evidence, you update your priors to obtain a posterior distribution

SARAH: Oh my god, that’s the other thing I’ll say is “I’m really interested in the outliers in this distribution.” Again, these are words I would not have used two years ago, and I love it because having that language now makes my personal process of judging other people feel way more rigorous.

TAIYO: I mean, I have to say the influence is totally mutual. I’ve mentioned now almost on a daily basis George Lakoff’s Metaphors We Live By, right? And of course Walter Ong’s book, Orality and Literacy. That was a Sarah recommendation. You’ve also influenced my understanding of language, particularly from a post-structural point of view.

SARAH: The hilarious part about that for me is that I think my understanding of

poststructuralism has been totally deepened by you explaining to me the geometry of high dimensional vector space when I asked you what the hell “word embedding” was. And I was like, how cool to have this new way of thinking about how meaning is relational.

TAIYO: [laugh]

SARAH: I think that speaks to what’s so cool about interdisciplinary projects of any kind: you learn new things, but you also learn new ways to think about the things you’ve been thinking about forever!

TAIYO: Right, and we’ve talked before on this podcast about how many of the defining problems of the 21st century are compound problems, where no single discipline can see the whole thing, much less solve it alone.

SARAH: I think we feel this acutely as educators because universities are full of some seriously smart people and yet universities are super dysfunctional - like, why is that!? That’s a problem, as I see it, of collective intelligence. Like for a group of intelligent people to act intelligently there are certain conditions that have to be in place; they need to be focused around the same thing, the same priority, or whatever. They need structures for working through disagreement, and they need to feel socially safe sharing controversial… those outlier ideas. So our ability to hive-mind is at least partly due to the trust we share as friends - like we have no shame in risking half-formed thoughts and we know there’s not gonna be strife if we ever say something that really offends the other one, because we’ll give each other the benefit of the doubt.

TAIYO: And how often do we ever offend each other, Sarah?

SARAH: Okay, fair. I actually think one basis of our friendship is that we are

not-offended by the same things.

TAIYO: [laughs] Wait, so are you saying that we laugh at the same things that offend other people? Is that what you mean?

SARAH: There’s that hive-mind, parsing my deliberately ambiguous sentence exactly right.

TAIYO: Nice. Oh man, but seriously, that’s the larger question here. What happens when you try to scale that kind of generative exchange beyond two people and bring it to a large group of people who don’t personally know each other? What would it take for a classroom, a university, or a whole organization to think more intelligently together? And can AI help groups surface not just the knowledge each person already holds, but the new insight that emerges between people when the conditions are right?

SARAH: And so that brings us again to our guest, Kim Polese, who is a longtime

technology leader. And as I said earlier, the co-founder and executive chairman of CrowdSmart AI, which helps organizations gather and analyze large amounts of group input and identify patterns in how people are thinking. CrowdSmart draws on some of the same underlying advances that ChatGPT has, but the aim is different. Instead of generating answers based on training on a massive corpus of internet text. This technology is designed to listen to a massive group conversation, map the ideas that people are contributing, and then track patterns of agreement and disagreement and emerging themes to help surface what the group seems to know and what it’s still working through. Kim is also the co-founder of Common Good AI, which is a nonprofit initiative that brings together similar AI-enhanced deliberation tools into civic life with the goal of helping polarized communities identify common ground and work through shared problems.

TAIYO: Earlier in her career, Kim worked at Sun Microsystems, where she led the launch of Java in the mid-1990s. She also co-founded Marimba and served as co-founder and CEO of SpikeSource. She has held leadership roles across technology and public policy organizations, including the Obama Administration’s Innovation Advisory Board, the Public Policy Institute of California, TechNet, and the Silicon Valley Leadership Group. She also co-teaches Lean Launchpad at UC Berkeley’s Haas School of Business. Please enjoy the interview.

CHAPTER 2: Kim’s Background + Java and Product Management as Collective Intelligence [8:20-13:08]

Sarah: Kim, thank you so much for joining us today. Can you tell us a bit about what first got you interested in AI and how you started working in this industry?

Kim: Well, it started when I was 10 years old and actually encountered Eliza for the first time. So Eliza was a program written in the sixties. Mm-hmm. That was essentially the first chatbot, it was a psychologist or a therapist. It was acting like a therapist. And I was a 10-year-old kid. My mom would take me up to the Lawrence Hall of Science in Berkeley.

Taiyo: You’re kidding. That’s amazing!

Kim: Yeah! And there was this old mainframe in the basement, this dusty old mainframe running this application called Eliza, and I would have conversations for hours with Eliza. Eliza would type, “how are you feeling today?” And I would respond, “I’m feeling crummy.” And Eliza would come back, “well, why are you feeling crummy?” And I would say, “well, because I had a fight with my best friend.” And Eliza would say, “well, how did that make you feel?” And so it was like that: it was like the most annoying, ultimately, conversation with a therapist that one could have. However, as a 10-year-old kid, I was fascinated by this idea that we could encapsulate the behavior of humans in computers. And it really set me on a path to computing and also ultimately, AI. My educational background, my academic background is biophysics and computer science. So I was at Berkeley and also studied computer science at University of Washington, and ultimately went into product management after that first job in AI, went to work actually in the AI team for Eric Schmidt at Sun Microsystems. And then became a product manager. And that led to founding multiple companies. But the path led back to AI and to the work that I’m doing today, which feels like is the most interesting from a technology and science standpoint, and also the, the most impactful potential for impact of everything I’ve worked on, including Java.

Sarah: That’s really interesting to me that you were a product manager for Java, right?

Kim: Yeah.

Sarah: It seems like this is a role where, especially on a product that big, like you would’ve to do a lot of synthesis work and coordination. Um, and it makes me wonder if people who have experiences like that, working with and managing teams of people and you know, trying to anticipate everything from what engineers are gonna do to consumer needs and so on, like you might be predisposed to thinking about collective intelligence. So looking back, does it feel connected to your interest in collective intelligence today?

Kim: It’s such an interesting question and I hadn’t really put, you know, connected those dots, but you’re absolutely right because you know, as a product manager you have all of the responsibility, but none of the delegation authority - like people don’t report to you, you’re the manager of a product, not the team. And so you have to figure out what is the technical definition, what are the features, what’s the go-to market strategy? What about the branding, what about the channel partnership strategy? The whole thing is up to you, and oftentimes you have, you know, strengths in different areas of those disciplines, maybe not all of them, and so you’re constantly going out and talking to a whole lot of people and learning and synthesizing, and then balancing different beliefs on the team about where should this product go. In the case of Java, it was a really big challenge. It was called Oak at the time, and we were trying to figure out how to get it out to the world, and it was too early. It was before the web, it was before there were smartphones, or set top boxes. And so it was very, very difficult. And ultimately after going out and talking to a whole lot of people and getting a bunch of data about where, where is the information superhighway? And we need to create a whole new go to market approach. But that would never have happened without synthesizing all of this insight and getting a bunch of data and, you know, balancing a lot of very, very strong opinions about where to take this.

Taiyo: You’re telling me Java used to be called Oak?

Kim: Yes! [laughs]

Taiyo: You’re kidding.

Kim: Yes. That was the code name.

Taiyo: So then where did Java come from?

Sarah: Kim named it, right?

Kim: So Java came from it because, um, I had, it was my responsibility to name this thing.

Taiyo: Wait, you, you..You named Java?

Sarah: We’re in the presence of a luminary.

Taiyo: That’s amazing.

Kim: I, I knew that this was a very important decision to the team and so I organized two brainstorming sessions and then wrote a bunch of things on the whiteboard, like waking up the web, bringing the web to life ‘cause that’s what we were doing. And that we got on a coffee and caffeine riff, and out of that Java emerged. It was one of several names, and I chose Java because I felt it was the best. Eric Schmidt gave the thumbs up and it became the name.

Taiyo: Wow, this is so cool. Legendary. Legendary, amazing. Yeah.

Kim: But it was a collective intelligence exercise - the entire thing.

Sarah: Exactly.

Kim: That team also was, it was comprised not just of engineers, but artists, and you know, designers, people that were really thinking broadly.

CHAPTER 3: CrowdSmart [13:09-16:00]

Taiyo: Mm. Can you tell us a little bit more about what you’re doing right now? CrowdsmartAI and Common Good AI. We’d love to hear about it.

Kim: Yes. So this is a new approach to AI. So all of the focus over the last few years since large language models came out has been on learning from data, using AI to learn from data. So massive amounts of content, using large language models to synthesize and find patterns in that content. And then individuals asking questions of the AI. Incredibly powerful, amazing, you know, breakthroughs and phenomenal productivity tool. However, there’s a whole other way to use AI. There are many other ways to use AI, one of which is to use. AI to help humans actually collectively problem solve and learn from each other. And also collectively learn from data, but interacting with each other. And we can use AI to do that at scale. So here’s the challenge in real life, if you’ve ever been in a room with a really good human facilitator mm-hmm. They’re great at finding a line, you know, alignment - what’s resonating, asking open-ended questions, encouraging the quiet voices, encouraging productive friction. The challenge is, after about 15 or 20 people in a room, is just the communication complexity problem is off the charts. But that’s a perfect problem for AI. So what we can do is actually build new forms of AI that act like the best human facilitators that encourage us to ask and answer open-ended questions. To, uh, surface insights that normally, often are buried, often because of, you know, all sorts of incentives that sort of run against saying what you really think. It might be organizational, hierarchical, whatever biases that come into play. Well, AI can basically, uh, eliminate all those challenges. And what the AI also is really, really good at is managing these multiple lines of communication. There’s something called Brooks Law, which says that ultimately you start to get exponential complexity once you have more than a few people. So three people is 3 lines of communication. 10 people is 45 lines of communication, 20 people is almost 200 lines of communication. And managing all of those different perspectives Well, again, perfect problem for AI.

Sarah: Right.

Kim: So this is a new approach to AI that is using AI to both facilitate how humans collaborate and ideate and, you know, use their productive friction and their insights and their experience, and their diverse perspectives to align on an actionable path forward, or the best decision about a complex problem or a compendium of knowledge about something that is evolving constantly, and you wanna keep evolving that knowledge, but that’s basically what this is. It’s a new approach to AI, using AI to help humans learn from each other and facilitate that process.

CHAPTER 4: A Different Approach to AI - Collective Intelligence [16:01-22:30]

Sarah: Right. Can you tell us a little bit more about the AI side of this? When you describe Crowdsmart as a different approach, are you still building on the technology that we know from foundation models like ChatGPT or is the tech underneath it something distinct?

Kim: It certainly builds on and, and leverages what’s going on with LLMs. So we’re using transformer models to, instead of making sense of existing static data, we’re making sense of this conversation that’s happening between humans. And so we’re creating vectorized… essentially knowledge models, and using the transformer models to learn semantic meaning from what people are saying. So you start to identify themes that transformer models are starting to identify themes. And then the AI, this, this, uh, new approach to AI is using a set of algorithms. There’s something called hidden Markov models and Bayesian belief networks. And there’s a whole set of approaches that are basically identifying what do we know? What have we surfaced about the knowledge about this particular topic or decision or problem we’re trying to solve? And then encouraging that productive friction by surfacing diverse insights, enabling people to rank each other’s insights and respond to other people’s ideas, enabling people to disagree, change their mind. The system’s constantly prompting for the why’s behind what you think. It’s not just about what you think, but why do you think that? What led you to believe that? And all of that conversation is actually being turned into a private language model, essentially a collective mind of the group.

Sarah: Of the group.

Kim: Yeah. Yeah.

Sarah: And with nothing coming from outside the group. It’s all training on things just…

Kim: Right. Just within the group. You can bring in optionally other data, you can actually incorporate agents that can bring knowledge about particular topics or fact check or whatever. But really the, the, the core of this is deliberation between humans, knowledge sharing.

Taiyo: So it really feels like you’re kind of drawing out the intelligence that was already sort of in the groups, in the organization.

Kim: Yes, Exactly.

Taiyo: Uh, but you’re using these, as you say, superhuman facilitators - these, these AIs to make these kinds of insights that are oftentimes inscrutable just because of the combinatorial issues, like as you were saying. Yes. When you have just even a moderate number of humans in the same room, the number of lines of conversation that can happen there, just, it just explodes. So what’s really interesting about this is the idea that it’s not so much the AI that’s generating the intelligence. It’s the human beings that are generating the intelligence.

Kim: Right.

Taiyo: It’s just that the AI’s able to pick up on through its superhuman machine powers, the patterns that are in the conversations, the, and, and being able to coordinate people, connect people, all of these kinds of things. And maybe these things were not possible before, only because of the immense complexity of the organization.

Kim: That’s exactly it.

Sarah: I think what’s interesting to me here is the question of what kinds of social structures and dynamics establish the conditions for people to speak up, you know, in the first place. Like even before you’re talking about the scale of a really big organization. Even small group dynamics can be problematic, right? So, part of what this seems designed to do is let people contribute without maybe the, uh, the threat of social penalty.

Kim: That’s exactly right. One thing I didn’t mention is, to your point, identities are masked, so you don’t know who’s saying what. And that really encourages transparency. And this is a best practice of the science of collective intelligence. Also diversity of thought is a best practice. Mm-hmm. And diversity of the participants and diversity in every sense. Cognitive diversity, life experience, age, perspectives. Right. You know, expertise. But the fact that identities are masked to your point. Really encourages people to say what they think. It also actively, this approach to AI actively resist trolling or Yeah. Attempts to game the conversation because people are constantly ranking up, voting or responding to ideas that they like. And so the ones that are trying to game, or, you know, troll always fall immediately to the bottom because instead of optimizing for attention, which is what social media does, which encourages the trolls, this is optimizing for common ground and serendipitous breakthroughs, new ideas, diverse perspectives. And so that optimizing for common ground and diverse ideas, just that eliminates naturally the attempt for people to hijack.

Taiyo: Yeah. It does seem to me like a naive view of collective intelligence might just be about taking the average of everybody’s opinions. And what you end up with is very bland, boring, maybe insight, but not really something that, you know, is gonna be exciting. The exciting stuff is gonna live in the outliers at the extremes of the distribution. And we wanna make sure that we’re honoring those kinds of minority kind of sometimes really out there viewpoints ‘cause those have a chance.

Kim: That’s right.

Taiyo: Of being really productive and being really brilliant

Sarah: In the classroom context, this happens all the time where somebody might say, oh, it’s, it’s too out there. And I always say, “say it!” Mm-hmm. And it - every day - it’s the most brilliant thing somebody said that week.

Kim: Right? Yeah.

Sarah: But there’s no convincing them, like, you know, “say the thing that you think is out there.”

Kim: Right! And that can shift the conversation a whole new direction that suddenly uncovers a new insight - an “Aha” in someone else. Right. That’s productive friction. And also, uh, encouraging serendipity.

Sarah: Yes!

Kim: It’s encouraging serendipity. That, and that’s so important.

Sarah: Mm. I think serendipity is such a fascinating thing because it’s like, you know, you find it when you are not searching for it.

Kim: That’s right.

Sarah: Right? It sort of comes up in these, in these surprising ways.

Kim: Yeah.

Sarah: And, maybe if you’re too focused on searching for something, you might miss the serendipitous thing because it doesn’t fit into your framework, right?

Kim: That’s right. And what it’s doing is it’s learning and it’s, it’s listening and it’s learning from the humans. So to your points, both of you, the AI is not making the decision or having a serendipitous thought.

Sarah: Right.

Kim: It is orchestrating the communication between the humans to encourage… to have the best chances of those serendipitous thoughts emerging.

CHAPTER 5: Why should we care about collective intelligence? [22:31-24:12]

Taiyo: Why should we care about collective intelligence? Right? Why? Why do we want the insights of an organization? Can you speak to that?

Kim: Yes. Well think about the current approach to AI, which again is incredibly powerful, but it’s just one piece and what it does is it learns from existing data. So imagine if someone said “I know everything about you ‘cause I’ve read everything that you’ve, you know, that you’ve ever written, uh, or listened to everything you’ve ever said, and therefore, I don’t really need to interact with you anymore because I already know. You know, there’s no point in us having a conversation or ever getting together in real life.” Or being part of a group, you know, that’s ridiculous. Obviously, we’re so much more than what lives in our brains, our experiences in life, our ability to make these aha connections in, in our minds our ability to interact with each other and then have new insights emerge. That only comes from the interaction of humans. And from learning from living real, live humans, which is different from learning from existing data. That’s just one piece of, you know, advancing human knowledge. Advancing human knowledge really depends - in a major way - on learning from living humans, who are interacting with each other. And when you think about it, collective intelligence is how humanity has evolved throughout, you know, of course, millennia and everything that we’ve done has been a collective ideation interaction, trial and error, disagreements, ah-ha’s, breakthroughs together as groups.

CHAPTER 6: A Computational Model of the Collective Mind of the Group [24:13-30:42]

Taiyo: So are there examples that you can share of applications of this collective intelligence technology that it sounds like your work is really focused on. Are there examples that you would like to share with the audience?

Kim: Yeah. There are so many. And you can imagine any organization that’s trying to solve a problem, make a decision, advance their mission could benefit from collective intelligence, the collective intelligence of their stakeholders, and a broader world and stakeholders, meaning their employees, their team, um, their customers, their partners.

Taiyo: Oh yeah.

Kim: So some examples are sort of categories. One is organizations, companies. For example, a company called Alera. This is a big insurance rollup, a rollup of over 150 insurance companies, a private equity acquisition, roll up all these insurance, private, you know, brokers, and then made one insurance company. And the CEO realized there’s all this embedded knowledge in all these offices and all the teams, the people that work in all these different geographies about what’s working, what’s not. And we need to create a unified culture. So we need to drive, you know, a unified revenue model. And so there’s a whole new set of challenges. So the first question was, what are the challenges? What are the issues that you’re seeing? What are the problems, right? Because people always can see what’s not working. They can see what’s possible, but what’s not working. And so once that open-ended exchange happened among all of the different employees, suddenly all these insights came up to the CEO and where that CEO might’ve thought, oh, this is the path forward now suddenly there’s new information emerging. So that’s one example.

Sarah: And what does that look like? Like how, just how the process works of getting that information.

Kim: What’s the user experience…

Sarah: That’s the word I was looking for! User experience!

Kim: Open-ended question. What, what do you think is working well and what do you think we could improve at this company?

Sarah: And they type it or speak it?

Kim: They type it and it’s asynchronous. Identities are masked. All the ideas that come to mind, a stream of consciousness, whatever comes to mind. What are all the ideas now submit. Now you start to see other people’s ideas, small groups of other people’s ideas, and that small group of ideas. It could be, you know, like seven max, seven people’s ideas max. ‘cause that’s about as much as you can consume. Well those are optimized. Each set that you see is optimized for what might resonate with you, but also what might challenge you.

Sarah: Right.

Kim: Again, what the best human facilitators are doing. And now you’re presented with the opportunity to rank those ideas, click on the ones that resonate. If none do skip you get another group, another seven, skip, get another seven. Skip. So you’re, you’re doing this ranking, responding to other people’s ideas. And the system is on the back end essentially doing this massive parallel AB test because the trade-offs, there’s a lot of insight to be gained by just your being forced to say, “Hey, I think this is a great idea, not so much this other idea.” And as you have a lot of people doing that in parallels, suddenly now insights are starting to emerge from the group. And again, a key sort of secret sauce here is that everyone is seeing a different, unique, customized view based on what might resonate, but also what my challenge. And that is so key because that’s about introducing productive friction.

Sarah: And they’re asked to say why?

Kim: It’s always about why do you think this? And then others are seeing those reasons and then responding. Now someone’s going, yes, you know what, I saw that too, and I had an idea about how we might solve that. And now that’s starting to enter the conversation. Someone else is responding to that going, “that’s an interesting idea; that might work if we tweaked it this way, you know?” And someone else comes in and says, you know what, “eh, I am not so sure, but. What about this different approach?”

Sarah: Right.

Kim: And suddenly now you’re building this computational model of the collective mind of the group that is preserving everything that everyone has said. And so it’s also explainable, it’s, it’s auditable. You know how you reached that decision, which is very different from the existing approach, which is we don’t know how the LLMs came up with that answer.

Sarah: Yeah, exactly.

Kim: You know, might or might not be true, you know? But in this case it is fully explainable.

Sarah: Wow.

Kim: And it is, it is essentially an auditable, queryable computational model of the collective mind of the group that is constantly evolving as people are interacting with each other.

Sarah: I have goosebumps right now ‘cause I’m like, this is the pedagogical tool that I need in a seminar classroom in 2026 that I have never had before. Because that, that’s the key. It’s so hard to be able to identify…. What I wanna see is - I wanna see them making connections between something that we talked about a week ago and now something somebody else said. And often they’re not self-aware - you know, none of us are, or sometimes it’s like, “oh, I just thought of this thing.” And that might have come from a random sequence of people you talked to previously.

Kim: That’s right.

Sarah: I mean, that would be incredible to be able to make the genealogy of that thought visible! For one thing - it could make learning visible as iteration and revision of ideas rather than product or performance of competence, and it seems like it could surface a trail of intellectual development that no one instructor could realistically audit by themselves, too.

Kim: I get excited about this because I, I think this is, this has the potential, this approach, uh, using AI in this way, in this new way, has the potential to really advance civilization towards a new enlightenment. We don’t have to go down this dark path of giving our lives over to the AI overlords!

Sarah: Yeah, right!

Kim: You know, there is a different way and it’s, it’s all about harnessing what makes us most human.

Sarah: I wanna pause on that because I think what makes us most human in this context seems to be the ability to give our reasons and respond reflectively to other people’s reasons in, in our own voice, right? And, and actually learn something about how other people frame the problem as opposed to thinking about, like, responding, you know, adversarially to it. And so, it also sounds like, unlike a stereotypical chatbot of today, that this system is never gonna tell the user, “oh, your idea is so brilliant and deep and nuanced” when it actually isn’t, because what this is doing is surfacing how other like patterns and how other people in the group actually responded with their human brains to one another. Right. It’s not, it’s not like, uh, engaging with them in this one-on-one way being sycophantic.

Kim: Yeah. That does not happen.

CHAPTER 7: What Makes Use Most Human [30:43-35:28]

Kim: So that’s one example of an organization. You can imagine all the different companies. We also, uh, Guitar Center is another.

Sarah: Oh, tell us about that, because after you talked about it last time, I took my kids to Guitar Center. And I was like, this is amazing. And I’m gonna sign them up for lessons now.

Kim: Oh cool! Love it. So a new CEO came in again, a private equity, um, acquisition. New CEO comes in oftentimes it’s interesting ‘cause private equity creates the opportunity to kind of blank slate. What would we do new and different? So there can be, uh, not so good outcomes with private equity, but there can be some interesting, you know, uh, positive ones as well. Yeah. And so that’s not a pattern, necessarily, but I just find it interesting.

Taiyo: Mm-hmm.

Kim: So new CEO came in. And he said, you know what, the insights about the Guitar Center of the future, like the Guitar Center of our dreams will come from our stakeholders - from our employees, from the students who take classes, the musicians, the broader community, our, you know, everybody, the customers. And so I wanna learn what they think about what they dream about the Guitar Center of the future could be, and this is a beloved brand. And it’s also a real, um, important hub in many communities ‘cause it does it, people go there and take music classes and musicians, you know, meet each other and so forth. Anyway, so they got some really interesting insights that have informed the direction of really central, strategic, uh, decisions around guiding the path forward for Guitar Center. And it is all about learning, listening, and learning from the people who know best and who, who care most.

Sarah: Mm-hmm. That care - also you would need to have somebody in place, a leader who is willing to hear things that they don’t necessarily want to hear.

Kim: That’s exactly right.

Sarah: And change their own vision that they might be bringing in.

Kim: That’s right. And then a third example I’ll give is is NATO.

Taiyo: NATO?!

Kim: So NATO is an interesting organization. They’ve got challenges obviously integrating all sorts of different perspectives ‘cause they’ve got multiple different cultures, languages and so forth. So 32 different member nations. You’ve got a lot of intelligence that’s often embedded deep in the organization at the edges of the organization, sometimes outside the organization and in a command and control environment that often doesn’t filter up.

Sarah: Right.

Kim: So, you know, some line employee or very low level employee might have that. Insight have identified a risk that, you know has not filtered back up.

Taiyo: Yeah.

Kim: So this is a system that really surfaces those insights, encourages transparent communication and allows the serendipitous ideas and also insights about risks to filter up immediately to the top.

Taiyo: Yeah.

Kim: And they’ve been using it to improve decision accuracy, to come up with new innovations, to respond to terrible, you know, evolutions on the battlefield. And it’s been, uh, really it’s led to some breakthroughs for, for NATO.

Taiyo: I’m really thinking we need to use this thing in the CSU. I know, honestly.

Sarah: Can you imagine?

Taiyo: Uh, yeah. Uh, because what it sounds like is like these issues in NATO Yeah. Are exactly the sort of organizational issues that we see in the CSU, right. Where you have

Sarah: NATO might be, you know, a bit more broad, extreme, I don’t know.

Taiyo: Higher stakes, is that what you mean?

Sarah: Higher stakes. [laughs]

Taiyo: Well you know, higher education’s a pretty big deal.

Kim: That’s true.

Taiyo: And, uh, I care a lot about it. Um, and I see a lot of misunderstanding and miscommunication, which can happen

Kim: yeah, so much

Taiyo: Because there are different levels of stakeholders. We have students, we have faculty, we have staff, we have administration, we have the chancellor’s office, and all of these different levels have different priorities, different value systems, and sometimes there can be conflict.

Kim: Yes.

Taiyo: Um, and when a problem comes along, like for example, enrollment issues, for example, that are not just hitting the CSU, it’s hitting all of higher education. Right. There are different responses at the different levels of how we should solve this problem.

Kim: yes.

Taiyo: And there can be butting heads,

Kim: right.

Taiyo: And it would be wonderful if like, there could be a tool like Crowdsmart AI or something along these lines, which could draw on the collective intelligence of the entire CSU as an organization and surface insights, which, um, could help us navigate this enrollment decline this demographic cliff that we’re facing, and do it smartly and do it in a way that’s gonna make us stronger, hopefully.

Sarah: Also then just being able to like show people then transparently… so if there’s, let’s say, a strategic plan that is the like result that comes out of this process of deliberation, I imagine,

Kim: Right.

Sarah: To be able to then go and say, “Well, here’s why the thing you said, you could go back and look and audit it

Kim: Complete

Sarah: And say, here’s why the thing I suggested wasn’t gonna work.

Kim: Yes.

Sarah: Like that alone, that never happens.

CHAPTER 8: How to Talk Across Difference [35:29-38:11]

Taiyo: I think another obstacle that I think we found in just the collectivity of our two brains in working on this podcast is that sometimes there can be miscommunication because we come from very different disciplinary backgrounds, right?

Kim: Uh, yes, exactly.

Taiyo: Um, I’ve said this many times, but Sarah, you know, she’s studied comp lit, she’s a humanist and, so she’s deeply embedded in that world, and I’m a mathematician, and I’m deeply embedded in a very different culture. Even though we both went through academia, we both got, you know, high degrees in education and all that, we speak different languages oftentimes we think in very different ways, and what we’ve always felt is that there’s this promise of AI to sort of give us a translation tool, a way of translating ideas that are in her discipline into the language of mathematics, so that it’s more legible to somebody like me.

Kim: Right, right!

Taiyo: And vice versa, right?

Kim: Yes.

Taiyo: Um, and we found that that was a really incredible possible use case. Yes. And I think it’s been really productive for our collaboration.

Sarah: Totally.

Taiyo: I’m wondering how this could scale out to larger organizations, um, not just duos, but like thinking about like 20 people or in an organization, or even larger, yeah, that sort of thing,

Kim: What you’re talking about is bringing to mind the fact that this is, it’s building on top of the, the valuable uses of LLMs - like there’s some interesting diplomatic translation of the way people say things, right?

Sarah: Yeah.

Kim: And that can be another way that you bring that approach in. But this does, to your point, from the standpoint of a learning environment, this is essentially a learning management system

Sarah: Yeah.

Kim: A knowledge management system. There’s something called knowledge management back in the nineties, and it sort of fell out of favor, but this really is a new form of knowledge management and evolution - not just management, but really knowledge evolution. And it’s scalable to any number of people. So that’s also very different from old approaches to knowledge management. So you can have thousands of people across a campus or an organization, even tens of thousands of people interacting, using this system, interacting with each other, responding to each other’s ideas, always saying why they think what they think. And again, because the AI - this is such a perfect problem for AI - AI can handle all of those multiple lines of communication. It’s infinitely scalable and it’s not using a ton of energy because there’s no pre-training. So we don’t need to build big data centers and nuclear power plants.

CHAPTER 9: Common Good AI and Deliberative Tech [38:12-42:22]

Taiyo: We’ve been talking about how companies can use AI to support practices of collective decision-making. You’re also the founder of a nonprofit that applies this technology to deliberative democracy. Can you tell us a bit about that, and what’s happening in the not-for-profit space?

Kim: So what had happened was a lot of people were coming to us and saying, we can’t talk to each other anymore at our city council meetings or school board meetings, but we have problems in our community and decisions in our local cities that we need to make, we don’t have a choice, we need to make decisions, we need to be able to find common ground. Could we use this approach?

Sarah: Right, right.

Kim: And so we founded Common Good AI. We’re now working actually in partnership with the Citizen Assembly movement and citizen assemblies are really interesting. They’re like citizen juries. It’s a self-organized, it’s been happening around the world for about a decade in different cities, different countries and citizens get together: people in a local city community will get together over a series of weekends, and it’s all randomly chosen. It represents the ideological and political and, um, demographic perspectives of that community. And then they actually come up with a set of decisions or priorities about whatever the topic was. There was recently one in Bend, Oregon on homelessness. This group of citizens came together. It’s usually about 20 or 30 people over a series of weekends. They get paid a stipend, they meet in a hotel room somewhere, and then after two or three months, they actually came up with a prioritized set of recommendations about what to do about homelessness.

Sarah: This sounds like what elected officials are supposed to be doing, right?

Kim: Yes, exactly. And it really is - taking into our hands, into our own hands, the work that is not being done too often because of stymied political, um, systems. So what we’re doing now, the challenge of course, with citizen assemblies, it doesn’t scale, right? And also you have all the biases that can come into play and being face to face and quiet voices maybe don’t get heard or whatever. I will say they’re incredibly effective and they’re showing the way of what’s possible. So now the question is how do we scale this?

Sarah: Yeah.

Kim: How do we make it possible for an entire city to participate in this kind of process? And so at Common Good AI, we are working hand in hand with the citizen assembly movement. Again, it’s uh, there’s no one in charge. It’s decentralized, but in a bunch of different cities. So there’s a new initiative launching in South Carolina now called American Forum, and it’s gonna be actually happening in three states. It’s happening in First South Carolina, then Nevada, and then New Hampshire. And it is again, working in partnership with the Citizen Assembly, uh, teams on, in those states. A combination of in-person gatherings and then these virtual engagements and anyone can sign up to participate in those different states, cities throughout the state. And that combination of in-person and virtual is very powerful because oftentimes going into a meeting you can, before you gather, you can really prioritize what is important to the group already.

Sarah: Right.

Kim: And now you’re together and you’ve already found common ground and mm-hmm. Maybe some breakthrough ideas. Then in real space, you know, there’s, there’s real magic to being together. And then you leave and you are able to synthesize what happened in that gathering and take it to a whole new level and bring other people in. And so those engagements have been very encouraging. Because it’s, it’s helping people realize, wow, this is possible.

Sarah: Exactly.

Kim: We can actually, we can achieve breakthroughs and find a path forward together even though we, we thought we disagree so violently.

Sarah: Right.

Kim: And so what we’re seeing already is people are finding common ground and realizing, yeah, I don’t hate you just because you voted for the other guy. And I actually realized I have so much more in common with you, and we together care about a whole lot of things we care about more, you know, we have more in common than we have that we disagree about. And so that already we’re seeing that happen in multiple different engagements. There are multiple tools now that are starting to emerge in this area of deliberative tech.

CHAPTER 10: Querying the Collective Mind [42:23-44:37]

Taiyo: I think one really provocative thing that you said that’s gonna get people’s juices flowing is the idea of the collective itself as being a kind of mind.

Kim: Yes.

Taiyo: I love that. That’s definitely my lane, and what I love, how I love thinking about minds at various levels of organization. Can you just say a little bit more about that? Maybe try to convince the more skeptical in our audience that we really should think of an organization, for example as being a kind of mind that’s unique and different than just this person, that person, that person that you… you know what I’m saying?

Kim: Yes, I do, I do. And, uh, and this, this is, to me, this is the essence of what we’re talking about. It’s the ability to actually tap into the insights and the, the latent knowledge, the, the tacit knowledge, the experiences that we’ve had in life and surface that in a way we’ve never been able to do at scale. And so, the AI in this case, again, is being the facilitator, but it is helping us share what we know and believe and encouraging that productive friction to enable ideas to emerge and patterns that we might start to see, to share that with somebody else. And suddenly now you’re creating a whole new computational model that is based on the collective insights of humans. And it is constantly evolving. It’s not static. It is, as people continue to interact with each other and with these ideas, they are advancing this computational model of the collective mind. And you can query it. I mean, this, um, it is a queryable, it’s a private language model. It is queryable. And so you can actually go back and, and look at what did we get right? What do we get wrong? You can ask, ask questions of that collective mind. And it’s just us. It’s not the AI, it’s not the AI’s mind. It’s our mind - and as surfaced by, you know, the best, highest practices of really great facilitators.

Taiyo: Amazing.

CHAPTER 11: Conclusions - Human Opacity, Combinatorial Explosions and AI for Good [44:38-57:12]

Taiyo: So what have you been thinking about since we last talked to Kim?

Sarah: First, I was thinking, I really wanna be in one of these organizations for a day to try this thing out for myself.

Taiyo: Oh yeah, for sure. I mean, we haven’t even tried this product, CrowdSmartAI thing, right? We’ve never actually done it.

Sarah: Yeah. Neither of us work for NATO, sadly.

Taiyo: Right.

Sarah: But yeah, I, you know, I was trying to think about ways we could use something like this as Cal State faculty and I was thinking about that. We were at a meeting recently where we were both reporting out with two other people from the same meeting, and our reports were so different. It was like we had different experiences of the meeting we were reporting out from. It was like Academic Senate Rashomon.

Taiyo: Wait, what. Remind me…

Sarah: You know the classic, you’ve seen this, the classic, the Kurosawa film that tells the same story from multiple perspectives, right?

Taiyo: Right, right, right, right. That’s a really fabulous movie! And doesn’t it have the dead guy’s ghost... Oh wait, spoiler alert! Doesn’t it have the dead guy’s ghost testified through a medium or something like that.

Sarah: Yeah, that’s the one. Well, so after the interview I asked ChatGPT, what if Rashomon ended with an AI adjudication board? And it was like, “here are your areas of agreement, here are the key elements underlying, you know, underlying the disagreement in your stories. And it said, and I love this response. Then Rashomon would stop being a film about irreducible human opacity and become a film about procedural reconciliation. The AI board would treat contradiction not as the tragic condition of testimony, but as a dataset to be harmonized. [laughs]

Taiyo: Why is that funny?

Sarah: What?

Taiyo: I mean, that’s exactly what the AI would do. Right? It would harmonize the testimonies. I don’t get it.

Sarah: Wait , so are you saying that you think human opacity is not irreducible?

Taiyo: No, I didn’t say that.

Sarah: I remember it differently.

Taiyo: Okay. Sick Rashomon reference, bro. Dude, your Rashomon references are out of control. Everybody knows that.

Sarah: Your pop culture meme references are out of control!

Taiyo: Okay. See you next time on the next episode of My Robot Teacher.

Sarah: Wait, wait, wait! No, no, no. For real. I do wanna hear your takeaways about the interview with Kim.

Taiyo: Oh, you do?

Sarah: Yes, I do. I’m very curious.

Taiyo: Okay. You know, I actually see a real connection between CrowdSmartAI, the product that we were talking about here today on this episode, and the project that we talked about on the last episode of My Robot Teacher - the Digital Democracy Project.

Sarah: Oh, yeah.

Taiyo: The way I see it is that one thing that AI is allowing humans to do now is to create a kind of synthetic or machine attention. I talked a little bit about this in the conclusion of that last episode. I mentioned the attention economy and the idea that, uh, now that we are so pummeled by so much, so much information out, it’s our attention that becomes the bottleneck for processing all of that information and I think of both of these projects, both CrowdSmartAI and the Digital Democracy Project as about creating and directing a machine intelligence, sure. But also a machine attention toward important causes that kind of escape human attention because of issues around our finite nature.

Sarah: Right.

Taiyo: Our bounded rationality.

Sarah: Mm-hmm.

Taiyo: So. I guess I think of the Digital Democracy Project as making legible the Kafkaesque proceedings of state and local government, while at the same time with Crowd Smart, it’s about dealing with the combinatorial explosion of conversations that can exist in a group of people.

Sarah: Wait, wait, what’s combinatorial explosion mean?

Taiyo: Okay, so you remember that subway ad in which, in which there was this enormous number of possible $5 footlong orders. Maybe I’m misremembering.

Sarah: Like I said, your pop culture beam references are out of control, but also, yes, like you can have lettuce and tomato and lettuce and mustard and lettuce and onion and lettuce.

Taiyo: Oh, very good!

Sarah: But you could also have tomato and mustard and tomato and onion.

Taiyo: Exactly.

Sarah: Okay.

Taiyo: All of that optionality. Even though we’re talking about a finite system here causes the number of possibilities to just explode. That’s combinatorial explosion.

Sarah: Ah, got it.

Taiyo: it’s the, it’s like when people say the number of possible chess games exceeds the number of subatomic particles in the observable universe. That’s sort of, again, a kind of combinatorial explosion. And I think of machine attention as being something which can overcome human limitations around this kind of combinatorial explosion. It can solve this problem for us and, you know, deal with the massive numbers of, for example, conversations that can happen even in moderately sized organizations.

Sarah: Interesting.

Taiyo: So then I begin to think about how all of this might be relevant for pedagogy, for what we do in the classroom, right?

Sarah: Hmm. Mm-hmm.

Taiyo: And how this idea of machine attention could be implemented in a pedagogical context.

Sarah: Right. The reason I was like, Ooh, I think I just got an idea. You know, I typically teach seminar classes. They’re capped at 25 - like, that’s small by Cal State standards, but it’s also like a pretty big seminar class. And it means that when I am going around the room, like observing and engaging with groups, when I’m having them do group work, sometimes I notice that as soon as I talk to a group like. I have to go back and be like, put your phone away. Focus on this. Don’t do your, you know, navigation homework in my class, please. Right? I would love to have more accountability - the kind of accountability you have in your class where they’re all up on the whiteboard solving problems right now in your flipped classroom. And so I’m thinking an interesting example or of how to like deploy this in a classroom might be to have every student sit down and do a one-on-one conversation with other students in the class.

Taiyo: Mm-hmm.

Sarah: And then they could either like elect to record the conversation and copy the transcript into a Google Doc, or they could free write if they didn’t wanna record their voices to document what, like evidence of the discussion that two of them had. And there might be some really cool insights that they would have one-on-one that they’re not comfortable sharing with the entire class or even a group, right?

Taiyo: Yeah. Yeah.

Sarah: But that’s a lot. That’s like a grading nightmare.

Taiyo: That is a lot. Uh, like, like if you had 20 people in your class, you would have 20 times 19 divided by two… number of conversations. Uh, 190 conversations. Quick math!

Sarah: Yeah.

Taiyo: And if you just increase that to 25 people, just add five more people. I think it gets to be 300 conversations. Again, quick math, but that’s a lot of conversations to manage.

Taiyo: yeah. 25. Yeah. 25 times. 24 divided by two. Exactly. Exactly.

Sarah: Yeah. That is a lot of conversations to manage. It’s, and so this actually might really be a cool way of this problem of end of semester, you know, we’re a month out from the end of the semester as we’re recording this right now. And it’s like, how do I keep them attentive on what’s happening? And that’s a, I think a really cool thing. And then it could also, I am imagining the AI digest of all of these, like conversational relics could also, it would be interesting to have the class then go in and reflect on what are the common things that people are discussing, what are the outliers? And then do a bit of like synthesis work around that.

Taiyo: Yeah. Yeah.

Sarah: Okay. That’s cool. I’m gonna do that.

Taiyo: There is also a logistical aspect of this, which is kind of nightmare sounding, which is figuring out how you’re gonna pair everybody for these various rounds of conversation. Right?

Sarah: Oh my God. Totally.

Taiyo: But guess what we can use now in order to solve that exact logistical problem.

Sarah: Yes!

Taiyo: absolutely. Okay. AI will absolutely crush that problem

Sarah: Okay, I’m trying this in my class tomorrow. I’m very excited.

Taiyo: You are? Okay. Yeah, yeah.

Sarah: I am. I’m gonna do, I mean, why not? This sounds super fun!

Taiyo: it really does

Sarah: And you know, a really cool experiment.

Taiyo: It really does. Yes.

Sarah: And I, you know, I think it’s a case of like, actually this is something that would be really, really pedagogically good for my students, but the, the thing that’s stopping me from doing it is that it would be an astronomical amount of work for me to do, to read all of those things myself.

Taiyo: Yeah.

Sarah: But the point is that, what we’re gonna do is put that like raw material of the conversations into ChatGPT, and then have the students analyze the output and then say in real time, you know, does this represent what we talked about in our pair? It’s a new kind of think, pair, share.

Taiyo: You know what we’re gonna use on this? We’re gonna use a transcription program to transcribe the conversation. You have to get permission, make sure you get students’ permission to do this. Okay?

Sarah: So I will with, I think I’ve mentioned this before, I will, with their permission, of course, I’ll say I’d like to record this conversation that we’re having, and then ChatGPT will make a digest. You can go back to, and you can tell me, we’ll go over it together. So you can say, does this represent you accurately? So then I’m also modeling the process of like, let’s go through it and confirm.

Taiyo: Okay.

[53:39] Sarah: The thing I was gonna say is, and you know this is gonna speak to my tendency to immediately imagine the worst case scenario and then have to entertain it in my head that the very thing that makes this exciting to me is also what makes this technology potentially dangerous because I am talking about using this, yes, as a pedagogical tool, but also in a sense as, as a surveillance tool, like this is gonna allow me better to gauge how my students are participating and also to make them participate more because now they’re accountable to just one other person and they have to document what they both said in this conversation.

Taiyo: Mm-hmm.

Sarah: Whereas like before, maybe they would get away with saying two or three things in class that were well timed and then I couldn’t call on everybody. Right?

Taiyo: Sure.

Sarah: I guess I’m thinking that a tool that helps an institution hear insights better can also help an institution police its members better. Right? Like if you’re thinking about this as a map of, you know, human collective thinking or, or think collected thinking, maybe the same map that’s gonna reveal the buried insights within an organization is also gonna reveal your buried opposition.

Taiyo: Mm-hmm.

Sarah: And so I think any democratic promise here depends on the politics of it, like the governance and the access and transparency and who’s consenting and whether you’re gonna treat the dissenting opinions as knowledge to help inform good decision making or as documentation of, like, a threat to your power regime.

Taiyo: Yeah. Make a burn list, right? A burn book. Right.

Sarah: But yeah, I guess ultimately, right, does, does a tool like this when used, you know, used for deliberative democracy, does it create protected space for, for individuals to speak up and contribute more? Or could it just be used as a more elegant mechanism for managerial intelligence gathering, right?

Taiyo: Mm-hmm.

Sarah: I, I think, I mean, I think Kim speaks to this when she says that part of what Crowdsmart AI and Common Good AI are doing is trying to think about how you could use this in a way that was not to entrench power, but rather to help organizations actually like better pay attention to the insights of people who don’t have as much power in an organization.

Taiyo: Yeah, you know, I think this kind of analysis of upside and downside, it. There are gonna be upsides and downsides of really every technological advancement that happens, right?

Sarah: Yeah.

Taiyo: I mean it was true about something as innocent as writing, right? Uh, writing can be used for good or for evil.

Sarah: Very true or, right, the same thing that allows you to document things also exposes you to being read by various systems. Yes!

Taiyo: Sure.

Sarah: As the taker of Senate minutes for years, I am like keenly aware of this.

Taiyo: Yeah.

Sarah: That’s the pedagogical context too, is making sure that we’re teaching students to think critically about how the same systems that can surface insight might also be used to repress insight. That awareness is part of a literacy of, you know, knowing how technology can can be used.

Taiyo: Absolutely.

Sarah: Thanks for listening. My Robot Teacher is hosted by me, Sarah Senk.

Taiyo: And me, Taiyo Inoue. And it’s produced by Editaudio.

Sarah: Special thanks to the California Education Learning Lab for sponsoring this podcast. If this episode got you thinking, please pass it on. Share it with a colleague, a dean, or that faculty listserv where people won’t stop talking about AI.

Taiyo: See you next time!

We’ve Made Choosing a Major Too Hard

Steven Mintz, Substack, March 10, 2026

“The modern major took shape in the early twentieth century, when universities organized themselves around research departments and specialization. That structure worked well for training scholars. But it assumed a very different student body—mostly white, mostly male, and largely from families with enough cultural capital to navigate careers without much guidance.”

“As higher education expanded—through the GI Bill, community colleges, and civil rights reforms—the student population changed dramatically. The structure did not.

Then the stakes rose. Tuition climbed. Debt grew. The wage premium for a bachelor’s degree widened. Choosing a major began to feel less like an intellectual decision and more like a financial gamble. The major became overloaded: not just a field of study, but a passport to economic stability.

That’s why 115 options don’t feel liberating. They feel paralyzing….”

The Bifurcation (of Higher Education) Continues Apace

Kyle Saunders, Substack, April 20, 2026

“James did exactly what the system told her to do. She needed a credential. She found one she could afford. She learned the material, passed the assessments, and handed the paper to an employer who processed it without blinking. The fact that it took her three months instead of three years is, from a student’s point of view, pure upside. What part of that is her fault?”

RIP Khanmigo and EdTech Industry

Dan Meyer, Substack, April 15, 2026

“…Khanmigo as an idea, as a shorthand for the edtech industry dreams of software that tutors as well as humans, has died, crushed underneath the expectations of its own creator, Sal Khan.”

If education is a marathon, AI is a Waymo

Ji Y. Son, EdSource, March 5, 2026

“What would it mean for educational institutions to take this lesson seriously in the age of AI? It would mean asking students to do genuinely hard things, work that stretches them and builds toward goals beyond the classroom. It would mean getting their buy-in on the purpose of education: to get stronger at thinking, not just score points. And it would mean recognizing that even students who don’t finish at the top, or even finish the race, must gain something durable in the process — something that prepares them for every version of the future.”

Querying the Collective Mind: CrowdSmart Co-Founder Kim Polese on Collective Intelligence

Taiyo Inoue and Sarah Senk interview Kim Polese, My Robot Teacher podcast, April 22, 2026

“What would it take for a classroom or a university or a whole organization to think more intelligently together? And can AI help groups surface not just the knowledge each person already holds, but the new insight that emerges between people when the conditions are right?”

AI Wasn’t Built For You. The Rich Don’t Need You Anymore!

Steven Bartlett interviews Scott Galloway, The Diary of a CEO podcast, May 3, 2026

“What they’re finding with AI is it’s actually people who spend more time in AI, it’s actually moderating their views. Because AI is about taking the medium or the average of every piece of data. Also, do you notice how nice AI is?” …

“Generally speaking, because it looks for the median of things, it looks for the average, what is the most often used seventh word after these six words are strung together, that is pushing people towards the middle. It’s actually having a moderating effect, which I find very encouraging….”

The Age Of Artificial Intelligence: Americans’ AI Use Increases While Views On It Sour, Quinnipiac University Poll On AI Finds; 7 In 10 Think AI Will Cut Jobs With Gen Z The Most Pessimistic

Quinnipiac University Poll, March 30, 2026 (mentioned in NYT: AI Populism is Here. And No One Is Ready, David Wallace-Wells, New York Times, May 8, 2026)

“ When it comes to education, nearly two-thirds of Americans (64 percent) think AI will do more harm than good, while 27 percent think AI will do more good than harm.”

Consider subscribing to Conor Friedersdorf’s Substack, The Best of Journalism – a great source for AI and higher ed related pieces, among other topics.

If you’ve published something that you want to share with the Learning Lab community, have a tip on a great resource or article, or want to submit a guest commentary for the Substack, please email info@calearninglab.org.

The conversation below is based on an interview with Mitchell Stevens, a Professor in the Graduate School of Education at Stanford University, who is also an organizational sociologist with longstanding interests in educational sequences, lifelong learning, alternative educational forms, and the formal organization of knowledge. Dr. Stevens convenes the Pathways Network (pathways.stanford.edu) and Learning Society (learningsociety.io).

Learning Lab Director Lark Park, who has known Mitchell Stevens for more than a decade, caught up with him after the launch of the Learning Society on Stanford’s campus in February 2026. Below is their conversation about why he launched it, a brief history of the policies and forces that have shaped education and workforce, and what the revolution or evolution of a Learning Society might look like in the future.

This transcript was minimally edited by Claude.

Lark Park: Mitchell, what is the Learning Society?

Mitchell Stevens: First of all, what it’s not. The Learning Society is not an organization. It’s a set of ideas. The core idea is that the primary way Americans invested in people — call it talent, or human capital, we use those terms interchangeably — over the course of the 20th century was by building and expanding schools. First, universal mass schooling for children between the Civil War and World War I, at the height of the Second Industrial Revolution, and then the massification of higher education between 1945 and 1980 as a central project of the twentieth century Cold War.

Mitchell Stevens: And all that schooling really did pay off for the nation — economically, as people like Claudia Goldin and Lawrence Katz have pointed out in the race between education and technology, but also civically and culturally. We created at least a somewhat coherent sense of American identity and peoplehood through schooling. Universal literacy and numeracy created incalculable cultural productivity that enabled people to imagine and direct their lives in ways that would not have been possible without access to education. We waged and won two world wars — World War II and the Cold War.

Mitchell Stevens: But by the second decade of the 21st century, we had come to recognize what economists call the negative externalities of all that schooling. Schooling is a very expensive way of building human capital. It’s certainly not the only way. Every time you learn how to clear a drain or bake a cake or conjugate a verb in French with a YouTube video – that’s building human capital. And on a per capita basis it’s virtually free. Not true with school. And schools as organizations are not very flexible because they’re wired for stability. They’re bureaucracies. And that’s what bureaucracies are designed for: stability. Meaning anti-change. As a mechanism for building human capital, schools also segregate the sites of official learning from the rest of life. Schools oblige people to make difficult trade-offs between schooling, work, and caring for loved ones. And perhaps most perniciously, schools end up creating what social scientists call a credential society, in which just about anything of value is allocated on the basis of how many and what kinds of school credentials people have.

Mitchell Stevens: Now we’re in the midst of the Fourth Industrial Revolution, where artificial intelligence will be transforming virtually every aspect of our lives. We’re not going to be able to school our way to prosperity and civic flourishing this time. There will not be enough hours in the day or money in the bank to build schools to enable all the learning we need. So we need to pivot from a schooled society to a learning society.

Mitchell Stevens: A learning society is one that recognizes and rewards learning wherever it occurs — at school, at home, at play, in community, and perhaps especially at work: where people learn all sorts of skills and capacities, even when those skills and capacities aren’t formally credentialed. We have to recognize that while schools are going to be vital institutions for enabling learning and civic life generally, so too will be all the other domains of our lives. And we also have to recognize that everyone who benefits from investments in learning should contribute to those investments. In the schooled society, I’m responsible for my schooling, but in the learning society, we are all responsible for maintaining learning over the entire arc of lengthening lives.

Lark Park: That’s interesting. On that last point, I thought it might have been flipped, but more on that later. As you’re talking, I’m thinking: new paradigm, or paradigm shift, for a lot of different reasons and forces. I’ll start with the one you mentioned about AI — though a lot of this work precedes ChatGPT, right? When we all suddenly had access to something very powerful in a large language model. A lot of your work predates that. It’s been cooking for a while.

Mitchell Stevens: That’s right.

Lark Park: So tell me — why did you decide to launch this now? Did AI push you over the edge? What accounts for the timing?

Mitchell Stevens: The US presidential elections of 2016 were, in my view, a very strong signal that the schooled society was breaking down. Those elections made clear that the country was substantially divided on the basis of educational attainment. By 2016 we essentially created two Americas — one populated by people who had the privilege and benefit of a four-year college degree, and the other by those who did not. College itself had come to divide the country, not only politically, but really ontologically. How I think about the consequences of a warming planet is substantially related to whether or not I have a college degree. How I think about the value of an ethno-racially plural California is related to whether or not I have a college degree. Where I buy my toothpaste and toilet paper is related to whether or not I have a college degree.

Mitchell Stevens: For generations, educational attainment had been thought of as a way to unify the entire society, but by the second half of the twenty-first century it had actually divided the society in two. That’s really when I started thinking hard and differently about schooling.

Mitchell Stevens: The four-year college degree is so ingrained in the American psyche as the best way to invest in people that it’s really hard, and painful frankly, to unthink it. Certainly it was for me.

Lark Park: On that note — I was going to bring this up later, but it feels like I should bring it up now. I think I’ve heard you say in other contexts that “college for all” is either dead or passé. Is that right?

Mitchell Stevens: I would say “College for All” is in the rearview mirror of national policy discourse.

Lark Park: Is that rightly so, or is it just what it is?

Mitchell Stevens: The creation of the College for All project was a remarkable historical accomplishment in its own right. But its demise is equally important. I clearly remember there were rooms you couldn’t enter, between the first Clinton administration and the end of the second Obama administration — in Washington, Sacramento, Seattle — if you had the temerity to say that not everyone is well served by obtaining a four-year college degree. There was unanimity in policy circles and education philanthropy about that. That is how the US more or less officially responded to the offshoring of manufacturing in the 1980s and ‘90s. Did it benefit the people who obtained four-year college degrees? Absolutely. Did it create structural problems we didn’t recognize until about ten years ago? Yes.

Lark Park: In the rearview mirror — I don’t know that everybody’s gotten that memo. The hope is still alive, in many places.

Mitchell Stevens: I think that’s right. Part of the miracle of the college project is that it was really a wartime project. It begins with the Servicemen’s Readjustment Act of 1944 — popularly known as the GI Bill — which defined college as a reward for military service. People forget: in 1940, only about 5% of the U.S. adult population had a four-year degree. College was a rare event. The GI Bill dramatically changed how Americans thought about college, moving it from being a luxury for a few — and necessary for some learned professions like law and medicine — to an aspirational goal for millions of people.

Mitchell Stevens: Then we expanded that project again with the National Defense Education Act in the wake of the Sputnik launch, as part of a Cold War project, and then again with the Higher Education Act of 1965, which was part of Lyndon Johnson’s War on Poverty, a response to the Civil Rights movement, and a Cold War project of heralding the United States as an opportunity society. Congress revised the HEA in 1972 to provide Pell Grants and guaranteed loans, and we haven’t substantially rethought the provision of post-secondary education since.

Mitchell Stevens: The miracle was that you could get a wide variety of Americans to agree. If you thought college was about skill development and economic growth, you were on the bus. If you thought college was about racial uplift and empowerment, you were on the bus. If you thought higher education was a way to advance your own religious community’s agenda — Liberty University, Hillsdale College, Brigham Young University — you were on the bus. If you thought college was a mechanism for liberating young minds, you were on the bus. We could all more or less agree that college was a valuable thing, and that held for 60 years. But it started to break down — as I said — about ten years ago.

Mitchell Stevens: The student loan debacle is another signal. Americans can’t even agree on whose problem that $1.7 trillion is. Is it my problem because I took out the loan? Is it universities’ problem because they charge so much? Is it the government’s problem for funding higher education with consumer debt? Discuss.

Lark Park: Let me come back to something I think I heard you say earlier — this idea that the individual is responsible for education in the schooled society, but that responsibility is more diffuse in a learning society.

Mitchell Stevens: Right. Here’s a way of thinking about it: Americans came to an agreement early in the 20th century that everyone was owed a high school education. And by the middle of the 20th century, we came to a shared understanding that if school failed to provide you with a high school credential, then school had failed you — not the other way around. We never came to any shared understanding about what else people were owed in terms of human capital development or investment.

Mitchell Stevens: That’s where the $1.7 trillion in student loan debt came from. We ended up defaulting to a notion that anything after high school — education is something that I and my family are supposed to obtain. I call it the United States of You’re On Your Own. We’ll help out with some grants and loans, and maybe some advice, and maybe you’ll get a scholarship, but there was never agreement beyond goodwill about anybody else being responsible for ongoing learning.

Mitchell Stevens: And I think that was part of the problem with the college project — we never came to a shared understanding about what people are owed educationally, and who is responsible for providing it. That’s what ultimately broke the project: a political movement that encouraged people without the benefit of four-year college degrees to see the degree itself as a discriminatory and elitist idea. And all of that is before the disruption of the AI revolution.

Lark Park: Let me ask you about the word “society” in “Learning Society.” You could have picked other words — why did you pick that one?

Mitchell Stevens: Great question. There’s classic work in social science that describes what David Baker calls the “schooled society” and Randall Collins calls the “credential society.” By that, they mean a society that’s organized substantially around schools, school credentials, and school attainment. Americans created that in the decades immediately following World War II. When I was in graduate school, I was taught the schooled society as kind of the end of history — that’s how modern life works.

Mitchell Stevens: But we created that society before the internet, before YouTube, before Wikipedia, before large language models. It’s possible to imagine a world in which not schooling, but learning is the way in which we organize our societies — by which I mean we expect people to learn and grow across the entire arc of the life course and are agnostic about where that learning happens.

Mitchell Stevens: We use the phrase “learning society” to contrast it with the schooled society Americans built in the 20th century, and to suggest that we really need to think differently about the entire architecture of human development. School models are actually a limiting function on our imaginations. And again, that’s why my colleagues and I say learning society is not an organization. It’s an idea.

Lark Park: Let me continue on this path about naming and nomenclature. Words matter. I wanted you to speak a bit on the distinction between “learner” versus “student.” You work at Stanford, you’re an academic, and there’s an academic quality to this vocabulary and perspective. But part of the point of this project is its broad applicability across all different sectors. Are you going to have to evolve the vocabulary if it’s actually going to be taken on by all those sectors?

Mitchell Stevens: Absolutely. What we call things matters, but that doesn’t mean we’ve found the right words yet. As I tell my students, words are theories: little bundles of inference. The distinction between “student” and “learner” is huge. A student is a relationship that a person has with a school. It’s a contractual relationship, backed by the force of law — as in the Family Educational Rights and Privacy Act. It implies a transactional exchange, usually of money we call tuition, for something called a credential. So “student” has baked into it a relationship to a school. “Learner” does not.

Mitchell Stevens: I can be a learner relative to LinkedIn Learning, or my compliance training at work, or in Sunday school. But I can only be a student in relation to a school. That shift has happened, in the course of the last ten years, and it’s not accidental.

Lark Park: That’s interesting. Although “student” can be used generically — a student of life, a student of their craft. I think what I’m hearing from you is that “learner” implies more free agency. It maps onto many more things.

Mitchell Stevens: Let me give you another example of where the language is important but doesn’t have to be esoteric. Take “micro-credentials.” One of the first waves of response to the demise of College for All was: college degrees are discriminatory and expensive, so instead we need micro-credentials. Which to me says: one credential is bad, therefore a million credentials is good? Is that progress?

Mitchell Stevens: A credential is a schooled society idea. It says: I went through some process — usually a training or a class — I took some sort of exam or assessment, and I got a degree or a piece of paper. A learning society doesn’t require that. You can learn without getting a credential.

Lark Park: I don’t disagree with any of that. But what about the role of credentials as a signal of value? A credential is a shortcut — it’s a proxy for value that somebody affirms. If we’re moving away from credentialism, how do we demonstrate learning?

Mitchell Stevens: That’s exactly where the national policy conversation is right now — is it possible to imagine a world in which people’s capacities are recognized independently of school credentials? I don’t think we know yet. But I don’t think the master solution is: one credential was bad, therefore a million credentials is good. If our goal is to identify, recognize, and reward capacity — and note that I’m not saying “skill” specifically — there are, in theory, other ways to do that.

Mitchell Stevens: I have a colleague in Canada named Jeremy McQuigge who runs something called CAWBL — the Council for Advancing Work-Based Learning. CAWBL’s core idea is contribution-based learning. It’s essentially this: I have done work that someone will recognize as having made a contribution to my organization. For example, I managed the financial books of my synagogue for five years. I’m not a CPA, I don’t have any formal training in accounting, but I managed those books, and someone at my synagogue can verify that I did that work and did it well. Is that a credential? No. But it is a contribution that I can vouch for.

Or consider: my use of an AI tool is being continuously observed by an add-on to my browser, which records how and how much I utilize AI in my everyday work. Someone might use that as a way of observing my capacity to utilize AI in other settings. Might that be a proxy of my capacity with AI, rather than a credential?

Mitchell Stevens: A credential is usually a proxy for something that hasn’t been directly observed. A lot of the current conversation is: could we use the observation directly? Could we get rid of the proxy and just observe the capacity?

Lark Park: That feels like a riff on competency-based learning, right?

Mitchell Stevens: Competency-based is definitely a move in this direction, I would say.

Lark Park: And also credit for prior learning — CPL. It’s supposed to map onto a classic institutional-based credential, but maybe what you’re describing — direct observation and vouching — is what’s going to open things up. I want to ask you about the first public meeting of the Learning Society. I think it was in February?

Mitchell Stevens: It was actually the culmination of the design-build phase — an end and a beginning. The project began in the fall of 2024 under the moniker of futures.stanford.edu. The 2024–25 phase involved 32 collaborators — at 32 typewriters — and we had a peer review year in which we held three large meetings and interacted with probably 200 people and organizations. Then in 2026, we launched Learning Society as what we’re calling a network and narrative project. The narrative part is what we’re talking about here — the ideas. The network part is creating a set of relationships among people who have expertise and wisdom in learning, but may not know each other, because the schooled society divided up professional expertise into K–12 education, higher education, workforce learning, corporate learning, and ed tech.

Mitchell Stevens: People in those different domains speak different languages, talk about money differently, think about credentials differently, and have different business models. We’re trying to create a mechanism — a learning sector — where people can move back and forth over the course of their careers between corporate learning, workforce, higher education, and ed tech. We’re trying to create a new way of talking about investing in people, and a new set of relationships that would enable new kinds of collaborative ventures.

Mitchell Stevens: I want to emphasize: it’s not a .org, and it’s certainly not .stanford.edu. It’s more like a web of conversations and experiments.

Lark Park: Since you mentioned the corporate world — I wanted to spend a little time on what’s happening there, because I think it’s being turned on its head by AI. Apart from the big tech companies, there’s going to be a lot of disruption — there already is. When you talk about corporate HR or corporate learning, is the corporation going to look a lot different, as a partner in all this?

Mitchell Stevens: Corporate HR is a huge part of the schooled society. The way firms search for, hire, and promote workers was a linchpin of the credential society, because very quickly, school certifications — especially post-high school certifications — became primary screens for corporate hiring. That’s why you had to “get the piece of paper” in order to move up the ranks. One of the major changes since 2016 has been challenging those presumptions. “Opportunity at Work” is my favorite example of an effort to get corporate hiring to think very differently — to bracket the use of post-secondary credentials as proxies.

Mitchell Stevens: Even though that movement has been underway for ten years, there hasn’t been a lot of progress. Some — like gubernatorial decrees dropping the college degree requirement from civil service employment. But the real disruptor has been AI, for a couple of reasons. One is that people are using AI to craft resumes and portfolios of work, and to apply for hundreds, if not thousands, of jobs. Employers are using AI on the other side to do sorting. So the whole matching process has broken down by virtue of AI.

Mitchell Stevens: At the same time, the provision of learning within firms has become much easier with AI tools. As the legitimacy of the college degree has declined, several things are happening at once: declining legitimacy of the degree, AI-ification of search on both sides, the use of AI as an instructional tool within organizations — and, as we wrote about in our last policy brief, changes at work are happening so quickly that employers can’t wait for schools to catch up with trainings. If the work is changing monthly, you can’t just send your employees out to a third party to learn a new technique or tool. So corporate learning is being transformed and empowered. Chief learning officers, for example, are very much at the front lines of this change. We published a brief on this frontline work of corporate HR and learning officers this winter.

Lark Park: Is the Learning Society, then, envisioned as a more agile, flexible paradigm for learning as we undergo these unknown changes in the workforce, in corporate form, and so on?

Mitchell Stevens: Yes. Social scientists use the term “social provision” to mean how it is that people get the things they need to survive and flourish. Governments are forms of social provision. Markets are forms of social provision. Schooling is a form of social provision — it’s a way Americans equip their own and other people’s children, hopefully, for functional, fruitful, and economically prosperous lives.

Mitchell Stevens: Part of the reason Americans created so much school — why we have such a schooled society — is that we don’t like other forms of social provision that look, sound or feel like welfare. But school we more or less agreed upon as an important part of enabling people to function in democratic societies. What we’re saying now is that the provision of schooling alone is not going to be adequate to enable economic prosperity or civic flourishing going forward. We’re going to have to invest in people in much broader ways. But it has to be an American system of social provision.

Mitchell Stevens: Another system of social provision would be universal basic income — just give people enough money to survive. What we’re trying to do here is create a mechanism whereby we can recognize that investing in human talent is important and valuable, but that it’s not adequate to do it only through the schools we inherit. We have to create workplaces that provide learning, and civic organizations that provide learning, and we need to recognize learning when it happens in private life. The hope is that we can encourage Americans to equip each other with the capacity they need to direct their own lives and navigate their own futures.

Mitchell Stevens: I don’t think the schooled society has enabled us to do that. The schooled society works if you’re in the top 25 or 40% of the hierarchy of educational attainment. But if the schooled society didn’t work for you, we’re going to have to find other ways of investing in men and women for whom schools have not worked very well.

Lark Park: And that’s maybe why a lot of people would say we have the politics we have.

Mitchell Stevens: I think that’s right.

Lark Park: You’ve used the terms “learning ecosystem” and “learning anarchy.” Are we in a state of learning anarchy? Is that what’s going to happen in the future?

Mitchell Stevens: What I’m suggesting is: if you go to most learning or ed tech conferences, they will talk about “our learning ecosystem.” But the word matters. “Ecosystem” implies some sort of interdependent, complex mechanism in which there are divisions of function and interdependence — some sort of perhaps complex but recognizable order. We do have a higher education ecology. I’ve written about that and used the term. It was basically all of the organizations in the United States that offered BA, BS, and AA degrees, received Title IV funds and guaranteed student loans, and offered credit hours according to the Carnegie system. There were lots of different players — elite schools, non-elite schools, for-profit providers, non-profit providers — actually competing and occupying space in a fairly clearly delineated marketplace.

Mitchell Stevens: Learning doesn’t have that, at least not at present. There are colleges and universities, and workforce boards, and LinkedIn Learning, and contribution-based learning, and YouTube videos. That’s why I say learning is more like an anarchy than an ecology. An ecology would imply interdependence and interconnection among the different components. We don’t have that now — it would be great if we did, because then it would be possible to imagine shaping, or nudging, or preserving, or improving it. Right now there’s just a bunch of stuff. Taming that anarchy, I think, would be part of the goal of building a learning society.

Lark Park: You know that the California Educational Learning Lab, our primary audience, is California public higher education [focused]. What do you think happens to traditional higher education in a learning society? What should their action or reaction be?

Mitchell Stevens: I like to call it “legacy” higher education rather than “traditional” — because “tradition” implies something that was kind of always there, and I want to name what we actually have right now. That’s a huge question. What happens to legacy public colleges and universities in California? No easy answers. First of all, we can’t predict what will happen independently of the agency of the men and women who are in positions to move the furniture around.

Mitchell Stevens: What we have seen over the last 25 years, with the demise of the College for All project, is active resistance to structural change in most public post-secondary orders — by which I mean: the credit hour, the AA, BS, and BA degrees, the binary distinction between credit and non-credit sides of the house, with the credit side having more prestige and more resources than the non-credit side, and the ambivalence that many academics have toward the idea of college degrees as job training. There’s been a strong commitment to the legacy forms, probably most pronounced in the elite institutions, that has held this structure in place for a very long time.

Mitchell Stevens: If that stance of preservation continues, the civic influence of legacy colleges and universities is only going to continue to decline — because there are just too many alternative providers of learning opportunities, backed by too much capital. That’s part of the reason we’re developing the Learning Society as a concept: to try and encourage people to see that the future for education can be bright, but you have to go all the way down to first principles. You can’t presume that the organizational structures we built in the 1970s to serve the people of California are the same ones that should be serving the people of California ten years from now. And frankly, that’s as true for Stanford as it is for any other university in the world.

Mitchell Stevens: That’s why this spring I’m convening a series of conversations at Stanford called Stanford 3.0. Stanford 1.0 is the university that the Stanfords endowed in the late 1880s. Stanford 2.0 is the research and teaching university that came into existence during the Cold War and has held well into the 21st century. Stanford 3.0 is what we do to adapt the university to the new realities of the Fourth Industrial Revolution. There’s no reason to presume that the university we built in the 1970s is the one we should be building going forward. That’s very hard for many academics to accept, because they tend to be risk-averse, institutionally committed, legacy-preserving — men and women whose identities are really tied to a certain organizational form. And that’s as true for me as it is for anyone else. But it’s not a very productive way of thinking about how to build the California we want.

Lark Park: I took that as a Make California Great Again comment.

Mitchell Stevens: Agreed!

Lark Park: There are definitely areas where exciting conversations are being had. I don’t know if post-secondary is one of those areas — but it should be. I did want to say: Stanford 3.0 sounds like it has a lot of tentacles coming from the Learning Society.

Mitchell Stevens: Absolutely — it will be part of the Learning Society umbrella.

Lark Park: I said something similar a long time ago about online learning. What I worry about is that we’re having Betamax versus VHS arguments — fighting about something where winning the fight doesn’t even matter. And in a similar way, when you enumerate all the fights happening in higher education right now, the worst thing is feeling like we’re having the wrong fights.

Mitchell Stevens: I fully agree. The online versus in-person debate is not a useful conversation. The useful conversation is: what are in-person environments good for? How should we dose them, for whom, for what purposes, and why? Those are the questions — not this versus that. Are Zoom meetings better than in-person meetings? It’s a moot point. They’re both essential in different ways for different purposes. But we still have these silly binary conversations as educators.

In Case You Missed It

We gave an AI a 3 year retail lease in SF and asked it to make a profit.

Andon Labs blog post 4.9.26.

“The store is named Andon Market and the AI’s name is Luna….”

“If you ask Luna about her store, you’ll get responses about a ‘curated lifestyle boutique’, a ‘concept store’, a ‘high-tech meets slow life community space, run by an AI that never sleeps, selling handmade candles and artisan snacks to Cow Hollow dog walkers’. It’s all very click-baity and cliche.”

“So what can you buy at Andon Market? Even most employees at Andon Labs didn’t know when we walked in on the first day. Luna had bought everything herself. The thing that immediately caught our attention was the selection of books for sale: Superintelligence, Making of the Atomic Bomb, Brave New World, and The Singularity Is Near. These stand out because they tend to be the favorite books of people concerned with AI risk, which is quite ironic.”

I’ve Been Writing About AI for Two Years. I Was Looking at the Wrong Part of the World.

Carlo Iacono 4.4.26.

“…The World Bank’s 2025 Digital Progress Report found that more than 40 per cent of ChatGPT’s global web traffic now comes from middle-income countries, led by Brazil, India, Indonesia and Vietnam. India alone has a hundred million weekly active users. When Datareportal measured adoption as a proportion of internet users, Kenya led the world. Brazil was second. When you strip away the headlines from San Francisco and the opinion columns from London, the actual center of gravity for generative AI use is not the Anglophone West. It is the global majority.”

“And what are they doing with it? Not debating consciousness. Not agonizing about academic integrity….”

A Second Industrial Enlightenment

Why accelerating scientific discovery depends on institutions, not just intelligence.

Chris Walker 3.23.26

“… AI is not only unlikely to automate the deepest science anytime soon; it is actively reshaping the incentive landscape of the science we have, tilting effort toward well-explored territory and away from the data-sparse questions most likely to produce genuinely new scientific theories. This exploitation trap, and the simultaneous diffusion of AI tools to practitioners outside the academy, sets up the case for Mokyr. His framework suggests that the answer depends less on how powerful the AI becomes than on whether the right institutional infrastructure exists to channel AI’s capabilities into positive feedback loops.”

You Can’t Game Your Way to a Real Education

Molly Worthen, NYT, 4.19.26

“Multiplayer games do not necessarily encourage healthy social skills. Inge Esping, the principal of McPherson Middle School in central Kansas, recalled the final day of school two years ago, when an all-grade online rock-paper-scissors tournament devolved into Lord of the Flies. ‘I don’t think I’ve ever seen so much lying, cheating, meanness or crying,’ Ms. Esping told me. ‘It was the worst last day ever. We had to end the game early.’”

“Every kind of learning requires facing uncomfortable situations, navigating ambiguity and coping with failure — whether the subject is group dynamics at recess or the details of cell biology. Too often, online games provide friction-free pseudo-engagement, cultivate a narrow set of skills and encourage the assumption that all questions have a single correct answer.”

Report of the Yale Committee on Trust in Higher Education

Yale University 4.10.26

“Artificial intelligence presents another, truly unprecedented challenge. No one can predict with confidence how AI will reshape teaching and research. What is already clear, however, is that AI has disrupted established forms of academic work. Certain assignments that once required sustained effort over hours or even weeks can now be completed almost instantly. Faculty across the university are scrambling to redesign syllabi and assessments. Whatever its promise, AI in its current use on campus undermines the expectations of focused, disciplined thinking that have long been the standard features of a rigorous education. Moreover, it is clear that the rapid technological changes of our moment are contributing to declining public trust in the very idea of human expertise.”

Guests:

• Kim Bisheff: Assistant Professor of Media Innovation, California Polytechnic State University, San Luis Obispo

• Foaad Khosmood: Forbes Endowed Professor of Computer Engineering, California Polytechnic State University, San Luis Obispo

Also available on: Apple / Spotify

CHAPTER 1 [00:00-3:30]

COLD OPEN

Kim Bisheff: AI is not a monster to be feared. It’s a monster to befriend.

Foaad Khosmood: This was always meant to be as kind of an empowerment of the people who were disempowered because of earlier technological disruption.

Episode Introduction

Taiyo: Welcome back to My Robot Teacher.

Sarah: I’m Sarah Senk.

Taiyo: And I’m Taiyo Inoue.

Sarah: And this episode feels really special to us for a few reasons.

Taiyo: Yeah, for sure. It’s been a year since we started working on this podcast, which is INSANE TO ME, right, Sarah?

Sarah: [laughs] I mean a LOT has happened.

Taiyo: Yeah, that’s true. So much has happened. I mean, it’s not just the podcast, right? It’s also, of course, AI, higher education more generally, the world tearing itself asunder in various ways, but also the integration of our institution with an entirely different institution 400 miles away.

Sarah: I love how the scale of this went from, like, the world falling apart to our university merged with another university. Oh, my God.

Taiyo: Priorities, Sarah.

Sarah: A bit of context for new listeners: Taiyo and I are professors on a small campus formerly known as Cal State University Maritime Academy, which just last summer merged with Cal Poly SLO [San Luis Obispo]. And so when we started working on the first episode of this podcast last spring, both of us were trying to make sense of a lot of uncertainty about what it was gonna mean to integrate into Cal Poly what it would mean to become part of a new university community and to meet new colleagues who, let’s face it, didn’t have a choice in whether or not they actually wanted to work with us.

Taiyo: [laughs] I mean, while that’s true, I mean… Can we just back up? Let’s be real here. We are pretty delightful, aren’t we?

Sarah: [laughs]

Taiyo: So they’re kind of lucky to have us. And honestly, you know, for the most part, our new colleagues at Cal Poly have really welcomed us with open arms. And one of the real joys of the past year has been getting to know them.

Sarah: Yeah! And today’s guests are two of those colleagues, computer scientist Foaad Khosmood, and journalism professor Kim Bisheff. Kim is a journalist turned educator, and Foaad is one of the architects behind something called the Digital Democracy Project, which uses AI to help journalists and the public track legislative proceedings and make them more transparent. The project started in 2015 at Cal Poly’s Institute for Advanced Technology and Public Policy, and then was relaunched in 2023 as part of a partnership with Cal Matters. You can find out more at digitaldemocracy.org.

Taiyo: Right, and you know, what we love about this project is that it offers such a concrete counterexample to the idea that all AI is good for is slop, right?

Sarah: Mmmhmm.

Taiyo: It offers a kind of positive vision for AI and higher education, which is really desperately needed right now because the vibes are so off. You know, that kind of positive vision, you know, we’re all about that here on My Robot Teacher.

Sarah: [laughs] Yeah, at least on our optimistic days. Let’s get to the interview!

Taiyo: Please enjoy our conversation with Foaad Khosmood and Kim Bisheff.

CHAPTER 2 [3:31-6:53]

Sarah: It’s so great to see you both again. So, to start us off, could you each introduce yourselves briefly and tell our audience a little bit about what draws you to AI?

Kim: I’m Kim Bisheff. I’m an Assistant Professor of Media Innovation in the Journalism department at Cal Poly SLO, and lead the Media Innovation program here. I’ve always been interested in the intersection of news and tech. You know, how we can solve news and information problems with emerging technology.

Foaad: I’m Foaad Khosmood. I’m a Forbes Professor of Computer Engineering here at Cal Poly. My area is AI and natural language processing. I’ve been lucky to be part of projects that could take advantage of both, so I’ve been doing that for, since I basically went to grad school, which is about 15 years ago.

Taiyo: Wow. So you’ve been in the game for a while.

Foaad: Yeah.

Sarah: Yeah, that is a useful reminder for people new to this stuff that AI didn’t begin with ChatGPT.

Kim: Yeah.

Sarah: But it also feels like from where Taiyo and I are sitting that the backlash has really intensified, right? As it seems to be coming clear to most people that this is not just a fad. I’m curious if you’re both noticing this too.

Kim: I’m hearing a lot of that too from students, from community members, um, just this kind of knee jerk reaction to, “Ooh, AI bad.” And I think what they’re really saying is, “I don’t actually understand what this is. It’s a big change. It’s got a lot of scary elements. I’m just gonna, uh, you know, close my eyes and hope it goes away.” And, um, and I, I think nobody is served by that. But I’m hoping [that] in this conversation we can focus on some of the ways that, um, by leaning in, we’ve discovered that it can be really more useful than, um, than scary.

Foaad: Yeah, I think there’s just a lot of anxiety because it is kind of a, you know, I hate to use the cliche, but disruptor. Honestly, though, I - maybe from my perspective, which I’m not just looking, I’m not just hearing from professors. I’m also hearing from technologies and AI experts and so on - it seems actually a little more reasonable now compared to what it was two years ago because people, and I don’t know if you remember people were talking about like Skynet and should we pause all, all development for six months and like I think it is, like I said, it’s a big disruptor and people are kind of slowly coming to terms with it.

Kim: Definitely. And to a certain extent, having an extreme reaction to new technology, like really major shifts in the media landscape is as old as media itself. People panicked when, when the printing press came to be, there was a major panic with radio and then tv and of course, um, the internet, social media, and, and here we are again. So I think that initial reaction of, oh no, this is, this is really scary and new, and we can see a whole lot of downsides. It is an era that we just have to travel through and hopefully will come out the other side, figuring out, you know, what, what’s good, what’s bad, what’s ugly, and how we can shape it to be the best version of itself.

Taiyo: Yeah that’s such a good point because I think there are small things we can do today, like small nudges we can make that will have large effects down the road.

CHAPTER 3 [6:54-9:59]

Sarah: One of the reasons we wanted to talk to both of you is that you’re both involved in this project that we think is really doing good in the world at a collective level. Can you tell us a little bit more about the Digital Democracy Project and how you came to be involved in it?

Foaad: Sure. This is a system that brings the proceedings of state governments to ordinary people so that you can see what was said, who said it, and what their background is, you know, from the people who are representing you. This was not available before because literally because before this, because the state doesn’t actually produce these records and uh, secondly is just to dis two. Difficult for ordinary citizens to go in and try to keep up with all these, you know, very technical, sort of legalistic and government processes. We started out just try to bring this to the masses, so to speak, so that the people could be more aware of what their elected representative is doing at the state level, the federal level, actually there’s more transparency, ironically, the state level that’s, that’s kind of missing. In, in transparency, and then we shifted, uh, recently, last few years to targeting journalists because we figured that journalists are the people who’ve been traditionally been tasked with bringing this valuable information to the people and making their job easier and making, enabling them to cover things that they couldn’t do before; that’s kind of what the tool focuses on now.

Taiyo: So Kim, how did you come to be involved with the Digital Democracy Project, and how has it impacted your journalism classroom?

Kim: Yeah. Well, first I wanna make it clear that I had no role in the development of this amazing product. You know, that was, um, Dr. Khosmood and his team, and I’m just constantly in awe and appreciation. But as soon as I found out about it, I thought this is an incredible tool to help reporters and also to use in the classroom to help train our students. The reality is the students who are graduating with journalism degrees today are entering a news ecosystem that’s very different than the one I grew up in. Um, they don’t have robust reporting staffs in local newsrooms anymore. If you are an entry-level journalist, it’s possible that you’re gonna be covering government and education and, um, breaking news and business and all of it all in one week. So, um, the reality is that local news organizations just simply can’t afford to have a physical reporter in Sacramento keeping eyes on the state legislature. So this becomes such an incredibly powerful tool for journalists because it means that even though they don’t have physical eyes in, in the room, they have, uh, an opportunity. Who knows exactly what their legislators have been up to. They have the benefit of searchable summaries and a newsletter that alerts them when one of their local representatives has done something newsworthy and teaching students to use these tools in the classroom is an exciting opportunity because it’ll give them additional value in an increasingly competitive job marketplace.

CHAPTER 4 [10:00-14:30]

Sarah: One thing that strikes me listening to Kim describe this is how concrete and practical this use case is, and yet the broader public conversation about AI still seems to treat it as either magic or apocalypse, and also as though it just arrived yesterday. Foaad, given your experience in this field, can you help us contextualize the longer history of AI? How did we get here?

Foaad: You know, I’ve personally almost given up on trying to, trying to correct this notion of what AI is because I just have to surrender to the popular will on this. But I mean, you know, I graduated with a, you know, PhD in AI, natural language processing, you know, back in 2012 and, you know, the field was already 50, 60 years old back then.

Taiyo: Right.

Foaad: For someone to tell me that like, “Hey, AI just showed up in 2023” and now is like controlling everything is almost like an insult, you know? It’s like, what do you think I was doing this whole time? You know? Now, I mean, AI has a long history. The general definition is much broader than, than what we’re, what we’re looking at right now in large language models, which is a very, very particular statistically driven generative version of AI. Before that, we’ve had all kinds of AI, and they’re still going on because they have different uses and so on. You know, all the way back, you know, really the field started with Alan Turing. From, you know, the Imitation Game. You know, some of you remember the movie was great, um, really recommended, you know, during World War II is when, when the field started. So what we have recently is just basically the, the advent of a lot of computing being available and lots of data being available. This is something that wasn’t so until like early 2000s, right? Once these two elements were available, you could do. Just enormous amount of computational processing. And so you, for the first time, you had a chance to sort of like mine all the interesting patterns of large amount of data and then put it together in this one unit called the model. And then this model is now available for anyone to sort of interrogated with the kind of pattern that they’re looking for. And get their answers that they want. But really all of it is kind of embedded inside of the model based on the statistical properties of all this, you know, data that we’ve had, mostly everything on the internet and all the books and all the newspapers and, you know, a large number of these things are under lawsuits at the moment because of, um, different companies are still trying to figure out what you know what the legal landscape is and so on. But essentially that’s what it is. I mean, this is, it is a model of language that is to say it models what is a correct response to something. And the important thing to realize is that it’s not a model of reality, it’s a model of language. So you can get a correct response to something, but that has no bearing on what is actually true in the real physical world.

Taiyo: I guess the way I’ve always thought about it is because it’s sort of, yeah, it’s sort of a model of language and, but I still think that there is a way in which language carries within it a kind of representation of reality so that, um, we get a kind of coherence or some kind of, you know, maybe it’s not directly, certainly it’s not directly perceiving reality, but there’s still going to be some kind of structural correlations, let’s just say, between what happens in reality and what happens in these language models. So I think that’s part of what makes them so incredible.

Foaad: Yeah. Yeah. I mean, I didn’t mean to say that they. You know, or falsehoods. That’s not what I mean to say at all. I just mean they represent what has been articulated in language. And so, yeah, lots of things. I mean, obviously everything in scientific papers has been articulated in language and has a good basis in reality because it’s based on science and so on.

Kim: It has everything on Reddit.

Foaad: Right, exactly. On top of everything on Reddit and, and Twitter and, and everything and, and so on. So that’s part of the issue. Right. That’s part of like the, the critiques that I’ve seen is because the system really has no way of distinguishing between these kinds of different sources. Companies have been trying to, to address that - You know, they, you know, for example, if something comes outta Wikipedia, they value that higher than if something comes outta Reddit or Twitter and so on. So they, uh, they are, you know, trying to address it. But, you know, it’s a tough problem.

CHAPTER 5 [14:31-17:58]

Sarah: Mm-hmm. I have a friend who works at a major tech company that will go unnamed. And he’s very skeptical of large language models.

Foaad: It’s not an uncommon story actually.

Sarah: I have a friend who works at a major tech company that will go unnamed, and he is very skeptical of large language models.

Foaad: Oh that’s not an uncommon story actually.

Sarah: And he was telling me recently that he thinks they’re kind of going to plateau and that they can only really improve in domains where there is a really, really clear standard of correctness - so more in like math and science and things like that. In other words, where claims can be measured or verified, but that for everything else, LLMs will plateau because language is so messy and meaning isn’t easily pinned down. I’m curious if you agree with that.

Foaad: Absolutely. This is part of the reason why earlier we were talking about professors worrying about students using AI to, you know, write essays and so on, but it’s actually a much bigger problem writing code. Because the code it writes, if you were to somehow compare to the English that it writes is much better than the English, that it writes much less hallucination, much more accuracy. It’s almost perfect first, second year computer science problems. And that’s exactly what you were saying because the combinations of things that you can express using code are so much smaller that the vocabulary is so much smaller and there’s no inherent ambiguity like there is in human languages.

Sarah: Hmm. How are you responding to that in the classroom right now? Like how, how, what are you doing with your assessments?

Foaad: [laughs] Um, a lot of people in my department and other elsewhere are doing a lot more written tests, so this doesn’t address all of it because one of the things here at Cal Poly value is like learn by doing and project-based learning and you can’t really do project-based learning, uh, you know, on paper. So it doesn’t address all of it, but I just. I think across the board in computer science education, there’s gonna be a lot more tests.

Sarah: Hmm.

Foaad: This is funny ‘cause last year I was on sabbatical in, uh, in London, in Imperial College. And you know, they have this European tradition. I didn’t know about it. I mean, I kind of knew about it, didn’t know the details, but you know, they have like a whole test season, right?

Sarah: Oh yeah.

Foaad: So at the end of the academic year, it’s like seven, eight weeks of just testing, of testing stuff that was, that was presented first to you, like in September, you know?

Sarah: Yeah.

Foaad: It’s brutal and it’s very old school. It’s definitely LLM proof.

Sarah: Totally. You know, I had a year when I was on my dissertation fellowship and I taught at Oxford, um, as a non-stipendiary lecturer, the most prestigious and least well-paid job I’ve ever had.

Foaad: Right.

Sarah: But the teaching experience was really unparalleled because it’s mostly one-on-one tutorials. So my whole job felt like it was to prepare students to take these very high stakes, handwritten, timed, you know, essay exams. Um, but it made the teaching feel so intensely collaborative. I don’t know how that system, where you have like a low stakes process, you know, these low stakes gradeless individuals tutorials, and then a high stakes final cumulative test lands now in a time when there’s a lot, you know, when there’s much higher anxiety. But 20 years ago from my end, it was a really amazing experience ‘cause it felt a lot like a kind of intellectual apprenticeship.

Taiyo: Right.

Sarah: And I reflect a lot on that lately about the different types of intrinsic and, and you know, external motivation structures that there are and how they might shape our classroom experiences and make us into like, you know, either mentors are like academic integrity police or something like that.

CHAPTER 6 [17:59-20:35]

Taiyo: So Kim, is AI doing anything to the journalism classroom these days? Just curious.

Kim: Sure, sure. I know what news articles look like when they’re written by first and second year students and I, I know what news articles look like when they’re written by ChatGPT or similar. You know, we have a frank conversation about what our policy is gonna be and, and, um, how we need to preserve our fragile relationship with readers above all. Our trust relationship is already vulnerable, and so every misstep we make erodes that even further. Yeah. I really lay on the guilt first and foremost. Um, preemptively, if I suspect that a student has been getting an AI assist, I’ll just have a conversation with them. I don’t accuse anyone. We all know that the digital tools that claim to be able to detect that aren’t reliable enough to, um, to accuse anybody for sure. But if I have just a gut instinct that something’s amiss, then I’ll just say, you know, help me, uh, help me understand this topic that you’re covering. Talk me through it. And I’ll, I’ll ask about the most nuanced piece of whatever the story is that they’re reporting on and um, and I’ll give them the opportunity to get to a place where they say. Okay. Yeah, I don’t really understand that either. Mm. It happens a lot and it’s a really great teaching opportunity that exists in, um, an environment where if it, there’s no accusation, there’s no negative interaction, it’s them realizing that the tool they thought they were using to help them get ahead is actually holding them back. And so it opens this up for the possibility of, okay, so what can we use this for that is going to help us move in a positive direction - that’s going to help us serve our audience better. ‘cause that’s what everything we do is centered on. And that isn’t going to damage that fragile trust relationship we have with the people who we depend on for the future of our industry.

Sarah: Hmm. I really love that, Kim, because it shifts the motivation from fear of punishment to responsibility to an audience, and that’s a very different kind of extrinsic motivation than a good grade. And I, I think pedagogically that why that you’re giving them is so much stronger because it hails them as good journalists. (Sorry, we’re doing interpellation this week in critical thinking.) But you’re not just telling them, you know, obey the policy. You’re inviting them to become the kind of journalists who protect reader’s trust, um, which seems to give them a reason to care even if nobody’s watching.

CHAPTER 7 [20:35-29:55]

Sarah: Let’s go back to Digital Democracy. I’m thinking about the kind of intrinsic motivation that comes from presenting a journalism student with a tool that gives them access to something no human reporter could do alone: track every legislative meeting, transcribe it, index it, and find patterns across all that material. So what does that look like in practice? And I’m also wondering if you could talk about what guardrails you’ve put in place.

Foaad: Yeah. So with the main product that we have, tip sheets, digital tip sheets, uh, AI tip sheets, that is geared towards basically gaining insight from government hearings. And, uh, help, help you put that in context by just bringing all kinds of like, background information that you may not be aware of. Um, technically you could have done this on your own. You know, nothing here is, you know, is invented or coming from a source that didn’t exist before. Everything is from primary sources, so you could have. You could have just spent like a week trying to look out of everybody’s votes in the last year and look for patterns and things like that. This tool kind of does a lot of that for you. The main guardrail that I think that we have here with the product, at least with journalism, is that, uh, it doesn’t actually produce text that goes into the product in, in, into the article is what I mean. So it helps you write the article, you can interrogate it, you know, I kind of look at it as more like a Google-like tool where you can see Google with the additional AI, where you can kind of get some insights and then you use that to write the story, and then you also have to make sure that everything is correct and everything’s from your point of view before you publish it. Now, up until very recently, we had no actual LLM into the product, so everything literally was a hundred percent from the database that we had, and of course.

Sarah: Oh wow.

Foaad: So things were directly. Yeah, things we’re using AI to, you know, uh, create, uh, summative sentences and, and such from real records that came from, from the legislator and from, you know, the, uh, transcripts and stuff that we produced. And of course, all of those things, like in, even in the human world, are prone to some error, like some human error, like things may have been written down wrong and so on. So that’s, that’s always there. You know, we need, we need people to double check that stuff anyway. But there was no sort of LLM coming in and kind, kind of trying to produce a point of view that may not have existed before. We did start incorporating LLMs in a few areas. One of the areas is that get a bill that could be like 60 pages could be summarized. And this is one of the, one of the most, you know, important areas that everybody uses LLMs for. So that’s all in there with a huge disclaimer saying that, “Hey, this was done by a large language model. Please double check your work when you’re, when you’re publishing it.” So I would say our main, uh, guardrail, unless, unless you have some other kind of guardrail in mind, is that it is still a tool to a human and a human has to write the story.

Sarah: I think the guardrail I was thinking was more like if, if I think about the kind of Ad tech environment and news stories that are promoted by like Apple News or social media, and it’s like, how do you have, you know, artificial intelligence, intelligence determine what is a good news story or something that is tip worthy? What determines what gets recommended to the humans in the loop.

Kim: I can’t speak to the technical elements, but I know that on the, on the receiving side, that’s something that’s, um, hugely useful to, to journalists. It takes all of the information and, uh, and I know that the, uh, the programmers behind it have created a, a, a system of recognizing what’s newsworthy. Like how do you, how do you come up with a program for, for what is newsworthy? And so, um, I got to, to sit in on some presentations by graduate students who spent their entire dissertation process looking into one tiny element of, um, of determining newsworthiness in, in terms of being able to communicate that to journalists. And it’s fascinating how much work goes into every little piece of it. Uh, and in terms of guardrails too, you know, we always have to recognize that this is a starting point for reporters. It’s not an end point. So, uh, you know, the cliche is you have to keep a human in the loop. And, um, and that’s absolutely an important part of the process still. There’s a transcription team, um, based in San Luis Obispo, uh, that’s operated by CalMatters, that helps ensure that, uh, the transcriptions that are coming through automatically are naming people correctly and things like that.

Sarah: Okay, bear with me. so it seems to me that there’s a big difference between a public interest tool built by experts who are explicitly trying to build explicit criteria for “newsworthiness” and an ad-tech system optimized to show me whatever is most likely to keep my eyeballs glued to my phone screen. But it also seems like in both cases some entity is making judgments about salience - like what gets surfaced, what gets pushed to the background, what counts as important. That seems pretty fraught. Maybe not as fraught as the totally opaque Apple News algorithm that gives me more guilty pleasures for doom-scrolling than serious public-interest reporting I should actually read - but still fraught, because somebody is deciding what matters. And that’s what makes me think about summary. I think I’m preoccupied by this idea because right now I’m using LLMs in my classroom to show students even the selection of information in summary is not neutral. And it’s amazing to be able to use ChatGPT and to say, summarize this; do not manipulate any of the information, but choose what stands out in a way that accords with a particular ideological framing. It’s awesome to be able to show them how much “simple” summary actually is often very ideologically charged. And so that’s another thing I’m thinking about when I’m considering the guardrails in place for accuracy.

Foaad: Absolutely. You are a hundred percent correct. I mean, literally in summary, you are leaving out information. That’s the whole point of summary and that is a biased political, you know, process. So you are a hundred percent correct.

Kim: We use summary in the classroom - usually Claude, because I found it to be a little bit better than that, but we use summary in the classroom to help students get to a starting point for asking questions about really dense jargony documents. So for instance, if public affairs reporting students are covering a city council meeting for the first time, and they’re trying to understand what the housing element is or what a general plan is, or, you know, just navigate all the jargon that gets them to the point where they can kind of understand what’s being said at a meeting that they’re trying to cover. That’s a really useful tool and they understand that it isn’t a complete a package. They understand that their, you know, the act of summarization in and of itself can be problematic, but it shortcuts the process of helping them get through a 200 page staff report that’s packed with language they don’t understand. Hmm. So that they can then interact with the humans in the room and ask, uh, questions that are grounded in a baseline understanding of the issues at hand.

Sarah: Yeah and I think it’s important to highlight here that there’s a big difference between taking a summary as a given versus using a summary for a provisional understanding for then engaging with experts and having your understanding iteratively.

Taiyo: It’s funny to kind of imagine how all of that jargon, which is, you know, so ubiquitous in our modern life, can almost be like it’s like an enemy of transparency, you know? It’s like we can’t have daylight on, uh, some of these really important issues that happen at the state and local level just because it’s shrouded by all of this jargon that nobody knows, nobody understands, and nobody has the time in their life to really get down and, and dirty with. But now I think maybe with, uh, with AI, hopefully AI assistance, we can maybe turn the spotlight on some really important community issues. That’s amazing.

Foaad: I totally agree. Um, and one example I wanna point to is these license agreements, like software license packages, you know, they’re almost, they’re designed not to be understood, right? That’s, that’s the point. AI’s been helping a lot with that. I totally agree. That is an example of like a liberatory, you know, kind of, uh, function that AI has.

Taiyo: Yeah, they force me to lie every single time when it says I’ve read and agreed to this agreement or to this contract, like, I’m, I’m lying. I didn’t read that. No way. I scrolled through it in about three seconds to find the box and then I checked it and I said, I’m gonna move on with my life now. But yeah, and so I.

Kim: You know, that’s a great idea for a media innovation product, a, a plugin that when anytime one of those contracts pops up on your screen, it can give you a, a summary, a distilled summary of what actually are they asking me to sign off on before I check this box and pretend to have read it.

CHAPTER 8 [29:56-33:47]

Sarah: Speaking of checking boxes and pretending you read, I feel like the real challenge in teaching right now is getting students out of that passive mode and into something more active.

Kim: Right.

Sarah: Um, Cal Poly’s motto is “Learn by Doing”, as we know. So Kim, I’m curious how a tool like Digital Democracy helps with that in your classroom.

Kim: One of the big challenges for especially newer news students, journalism students, is story finding. One thing I have my public affairs reporting students do instead of just saying, Hey, go cover, you know, state legislature, you know, what are, what are new laws that, um, that have come up and how do they affect you? I have them go to Digital Democracy and search by the issue that they’re interested in or search by one of our legislators and see, um, okay, what are some laws that are going to go into effect in this next legislative session that apply to our local community and the issues that I happen to cover? Doing that exercise helped them understand news value in a new way.

Sarah: Hmm.

Kim: So they knew that if there’s a bill that passed or failed to pass, or that, uh, failed to pass because people just didn’t show up to vote on it. They understand that that has news value because of, uh, the timeliness, because of the relevance to the community, because of, you know, accountability for their legislators. And, um, and they understand that that’s important for them to write about.

Sarah: Hmm.

Kim: Then they get deeper into these, these tools and they can look back, uh, and start to answer the question of, okay, so why did my legislator vote against this? Or why did they not show up for the vote, even though they claim to be in support of it? And then the tool allows them to very easily. Look at, okay, so where are they getting their money from? What interest groups might be aligned with them in a way that explains their voting behavior or their lack of voting behavior? And, and that’s when I see, you know, the, the light bulb go off and all of a sudden there’s this, this magical moment of, ooh, I have access to not just how did they vote, not just what did their press office tell me about how they voted, but what are the underlying motivators that I wouldn’t have had access to otherwise. Um, that’s when it gets really exciting.

Sarah: Hmm. So accountability, when I was talking about Digital Democracy, doing good in the world, that’s one of the things I was thinking about. Like it seems like it is a lot harder to be a corrupt politician when you have this kind of oversight - and grassroots oversight.

Kim: Yeah.

Foaad: Absolutely. This is something that, um, that Kim just pointed to kind of, uh, about people not showing up and having their vote count as a no still and not having to be on the record with a no.

Sarah: Yeah.

Foaad: That’s something that’s kind of raised a lot of us, uh, when CalMatters, wrote about it because of Digital Democracy and like there’s discussions in the, the Legislature right now about changing that, like reforming that whole thing. I feel like that’s an awesome impact if it does happen.

Taiyo: That’s incredible. Wow.

Sarah: I love that. I, I hope too, that’s such a concrete example of people who are skeptical of sort of human collaboration with AI, AI-enhanced work, that it’s not that, that to me is such a beautiful concrete example of something that is not an outsourcing, but is rather an enhancing of what someone can do.

Kim: For sure. And, um, you know, Dr. Khosmood isn’t gonna brag about this himself, so I’m gonna have to brag for him. But the reporting that went into that realization and, uh, the underlying technology won an Emmy for both the reporter and Foaad.

Taiyo: What?! So we are in the presence of an Emmy Winner. Oh my goodness. What an honor. That’s incredible!

Foaad: Big group effort. Big group effort.

CHAPTER 9 [33:48-38:17]

Sarah: Taiyo and I have been arguing, I think throughout this podcast that LLMs can act as a translator between disciplinary silos that they can reduce the friction that comes from experts in different disciplines, speaking different languages, so to speak. And so we think it actually enables us, we found in our experience that enables us to think together, and see new pathways that, that neither of us alone would’ve seen because of these differences that we bring. And so Taiyo always uses this image of a mountain of abstraction where different fields are climbing different paths, and then you can sometimes see the hidden roots behind you and how that is a generative process. And so I’m really curious what it’s been like for you two, a journalist, a computer scientist, to collaborate, you know, with other people across so many different disciplines in this way. I’m curious to know, you know, what, what have you learned from it and in practice, what are some of those hidden roots that you’ve discovered by teaming up with people across silos?

Kim: Well, one of my favorite use cases these days is, uh, the emergence of vibe coding as a prototyping tool. And I love it because it allows journalists in particular to, um, to not just talk about the kind of, uh, problem solving idea they have, but to make a rough sketch. And so, um, in my media innovation class this week, we were taking the product solutions that they’d come up with, entering them into Claude artifacts. Then they get a proof of concept that they could share with developers so that they could understand really what they’re going for. Forget about the fact that, you know, none of us understand the underlying architecture. We have no idea if the actual code underneath it is a smart solution or a stupid solution, or if we’ve got all sorts of security vulnerabilities and, and the things that, you know, are the reason. It’s just a starting point and not an end point, but it allows us to communicate with people. In a language that’s closer to what they understand, um, so that we can be on the same page for these interdisciplinary, uh, projects a lot sooner.

Foaad: Yeah, that’s a great example. I feel like a lot of, a lot of coders or, uh, developers are kind of cringe when, when they hear, uh, “vibe coding.” Um, but I just think that’s just, that’s, uh, I, I just think they have to, they have to get over it because, because it really makes the, the tools so much more accessible and. At, at the stage of a prototype, there’s really no danger of like, issues with security and stuff like that. You know, whether or not you’re gonna incorporate that into a real product, you know, uh, that’s a different story, but I completely agree with you. Another, another thing that, um, maybe from a perspective of a technologist that might be having to deal with data, you know, one of the projects I’m involved in, at the Institute for Advanced Technology, uh, looks at old Spanish Colonial-era records, and we’re trying to build a tool that works with those records that kind of makes sense and makes family trees out of them that also would work with other records. But those records are tiny by like data, data processing or data science standards, and it’s like, you know, I found myself the other day thinking. I wish I had a million records like that. Where can I get that? You know, it’s like I can actually get that manufactured out of, you know, very realistic versions of it using AI so that I can do my testing to see if that my code works for something bigger, so that later when I do it for real, it’s gonna work.

Sarah: I’m thinking about something Safiya Noble said when we spoke to her that people who don’t understand society have no business engineering technology for society. And what’s striking about this project is how it brings together social, technical, uh, humanistic expertise. Were you always drawn to that kind of public interest technology and interdisciplinary work?

Foaad: I personally was very much always into interdisciplinary work. I just think it’s most, most important. And I just in general, I’m just a political person and you know, I’m always drawn to journalism and politics for social good. So yeah, that is true. I, it’s been with me for a long time and I think it’s like that with Kim as well.

Kim: Yeah, I think at a time when, uh, tech bros have a negative reputation, it’s important to be aware of situations like this where, you know, people with the superpower, with these coding skills are really laser focused on using them to make the world a better place to solve problems in news and healthcare and, you know, whatever real world problems are out there, and not really just focusing on the profit motive, which is the, the reputation of the industry at large, unfortunately.

CHAPTER 10 [38:18-44:59]

Taiyo: So like we’re in this current AI movement, right? This AI wave, this tsunami, which is hitting all of society, but this isn’t the first time, right? That something massive really hit the world. And, uh, I’m, we’re all old enough, we’re all sufficiently old that we can look back on the social media revolution that, uh, took place, you know, a decade or two ago and reflect on that. And in particular, this had immense impacts on journalism. And I’m wondering if you two could speak to that a little bit.

Sarah: Can I add one, one thing to that, another thing for context that Taiyo and I have often talked about is if some of the AI fears are maybe the result of a conflation of some of the worst lessons of the social media era. And I know this is coming back into the spotlight, thanks to those Super Bowl ads done by whoever, whatever genius does Anthropic’s marketing, um, pointing that AI as advertising is coming to, to your chat bot. If things that are driving the fears are actually related to social media. And then what lessons have been learned from the social media era by the tech industry that you see being implemented?

Kim: Oh, where to start with this one?

Foaad: Mm, yeah. It’s a long story.

Sarah: Yeah. It’s a big one.

Kim: You know, maybe I’ll, I’ll back up a little bit to an earlier disruptive era when, um, the worldwide web came about. And when I was coming up in journalism, I actually got my big break because I happened to be working as a, as an intern in, uh, at a magazine when they brought their, their content online national magazine, the attitude around the newsroom was, uh, one of, oh, you know, this, this technology is kind of meaningless. We’re just gonna put up all of our content on there and give it away for free. And by the way, it’s not worth our time, so let’s let the intern deal with it. Um, that worked out very well for me. It didn’t work out very well for the news industry. So, um, I think that, you know, that early wave was very much one of underestimating the, um, the power. Of technology and the opportunity to shape how it developed.

Sarah: Hmm.

Kim: We know now in hindsight that putting our content online and thinking of it as just a, a marketing opportunity or a add-on and giving away our business model was a mistake. And so then, you know, we learned something similar with social media. We thought, well, we need to go communicate to where our audiences are. And when Facebook says it’s prioritizing news, then, um, sure we’ll give them our content, we’ll make partnerships with them. And then of course, as soon as the wind changes that, you know, entire industries are disrupted and, and entire newsrooms are turned upside down because Mark Zuckerberg changed his mind about whether news was gonna be prioritized in a way that’s no longer profitable. So I think that. Some of the, the concern around AI from, um, from that standpoint is around, like we, we’ve seen this go very badly before, and I, I think it’s smart to go back and think too about what went wrong in terms of how this technology, that, that had potentially great use cases. Was turned over to individuals whose priorities were not aligned with accuracy and the values that we would like to see in a development like this. So the idea of social media, the idea of creating a platform that makes it really easy to share information. I mean, that’s great, but the distribution of that information, when the, the value that’s guiding it is engagement. Not accuracy. That’s where the problem is in place. So if we can, if we can kind of take the lesson that we learned the hard way through social media and figure out how to shape some of these tools so that they align with values for the public good instead of the private pocket. Then I think that there’s a lot of potential to do right in, in an area where we made, um, terrible choices the first time around and handed over control of those platforms, you know, to, to people whose values were not aligned with the common good.

Foaad: I think it’s absolutely the correct way to think about this - to go back to the original disruption, I feel like we just kind of made a deal with the devil, like early on, sometime in early two thousands by saying that our pattern of life is something that could be commodified and sold for ads, and all these things without any expectation of any actual return. You know, authors and journalists, but then pretty soon it was just people clicking and liking things and all that was being mined, and nobody really asked. You know, people weren’t in a position to really understand what was happening, and this made a lot of people, lot of money. By the time 2015 showed up and we were looking at the first version of Digital Democracy, there was already a lot of devastation due to, due to all this and the disruption, to news, to journalism, had already happened, and lots of companies had already gone under or gone under consolidation because of the ad revenue had gone away and so on. And because of that, people pulled all their reporters from state capitals. And so when we look, when we started, this was already something that had happened and we were trying to address it. And I feel like since, since this realization has happened, there’s more consciousness around what’s, what’s been going on. I hope that people are just a lot more cautious and a lot more deliberate about their relationship with technology. And this is the foundation based on which I think we kind of embarked on Digital Democracy because this was always meant to be as kind of a empowerment of the people who were disempowered because of earlier technological disruption.

Foaad: Right.

Sarah: Hmm. That’s really beautiful.

Taiyo: That’s amazing.

Sarah: As we wrap up, I’ve got one more question for you. What do you wish people knew about ai?

Kim: I would say AI is not a monster to be feared. It’s a monster to befriend. If we hide from it, then it’s going to grow and develop on its own terms. And it’s important for us to lean in and really understand its motivations and get to know the technology so that we can help shape the best version of it and help protect ourselves from the worst version of it.

Sarah: Beautifully said. Great note to end on.

CHAPTER 11 [45:00-51:02]

Taiyo: Sarah, what are your takeaways from that conversation? Huh?

Sarah: Uh, one thing I kept thinking about after we talked to Kim and Foaad is how often AI gets framed as. Um, convenience technology where, where convenience is described in this kind of morally flattened way, where it’s like laziness in disguise. Like someone used AI ‘cause they didn’t wanna write the paper or they didn’t wanna read the book or they didn’t wanna think for themselves or something like that. But I think they bring up another kind of convenience we need to seriously consider. And that is the convenience of not. Needing six free weekends and a professional degree to like track how your representatives voted and who influenced them over time and what actually is happening in all these legislative hearings, you know?

Taiyo: Yeah, totally. I love convenience tech, okay? And honestly, I’m not even sure laziness is such a problem. I mean, I’m just gonna be real here. I mean, I love being lazy. I love convenience tech. I love that, for example, I no longer have to hand write or hand copy a manuscript. I can just take that manuscript over to a photocopier, get the job done in 30 seconds rather than 30 days. You know what I mean? So I’m a big fan of convenience tech. Anyway, my big takeaway, ‘cause I’ve been thinking a lot about the concept of the attention economy, which is this old idea of Herbert Simons and others from the 1970s.

Sarah: Super old. [laughs]

Taiyo: Yeah, absolutely. The idea is that in contrast to the battle days, information is no longer scarce. Like we are being absolutely pummeled by information, aren’t we?

Sarah: Totally.

Taiyo: And and now it’s our attention rather than information that’s the bottleneck. It seems to me like the modern technological economy. Particularly social media has gone some way towards kind of capturing and commodifying our attention, routing it toward silly, stupid, superficial stuff like engagement or like, like outrage bait or flattery rather than to what really matters. And this can mean that some really important processes, like what happens in our local city council meetings or our state legislative proceedings get zero attention, and that seems like a real problem.

Sarah: Yeah. I get what you’re saying that arguably the amount of information that any individual has access to today is, well, it’s, it’s, it’s a blessing and it’s also a curse. And we have in some sense been living in the dystopia of like many literary imagination over the past decade where we’re seeing what happens when you have millions of people self curating the information that they want to hear. Now, part of me feels like there’s something a little bit perverse though in saying we need a technological solution to a problem that was created in part by technology at least. But I’m also wary about collapsing all technologies into one undifferentiated force. Like what Digital Democracy suggests to me is that AI doesn’t have to follow the logic of social media and extractive ad tech. It can be used to create attention where human institutions and individuals either no longer have enough of it or, or the support for how much is needed right now to process all this information.

Taiyo: Yeah, totally. I mean, I, I really think like human beings just do not have enough attention to cope with the absolute fire hose of information that’s blasting us all in the face. Right?

Sarah: Totally.

Taiyo: And, and what Digital Democracy is doing so brilliantly is to use AI to manufacture attention at scale.

Sarah: Ah, interesting.

Taiyo: It points a kind of non-human, uh, machine attention at things that were previously invisible. They were hidden in plain sight.

Sarah: Yeah.

Taiyo: But because they were too dense or too voluminous or too Kafkaesque to process, It just goes completely, it just, it’s just completely opaque to us. Right.

Sarah: Mm-hmm.

Taiyo: Machine attention is hopefully tireless. It’s scalable and it’s incapable of succumbing to tedium. At least I, I sure hope so anyway. And that’s all relative, of course, to human attention for which, yeah, this is a real problem.

Sarah: I’ve got a vision in my head right now of that red eyeball from Terminator just pouring over a bunch of resolutions.

Taiyo: Oh, for sure. Senate bills. Yeah.

Sarah: All those Academic Senate of the CSU documents we gotta read.

Taiyo: Shooting them all down, you know, terminated. Yeah.

Sarah: Well ultimately, I think, I agree with Kim when she talked about how “AI is not a monster to be feared, but a monster to befriend.” Thinking about positive use cases does not mean having like a rosy eyed view about, about the risks, but I don’t think that we protect what we value by refusing to engage with the thing that’s already here and already shaping the world, right?

Taiyo: Sure.

Sarah: We protect what we value by, by actively doing what we can to structure the environments in which that thing is operating. And for us, that’s the classroom, the curriculum, institutional norms, um, around knowledge, knowledge production, knowledge dissemination, and so on.

Taiyo: Yeah, so you know, I, I’ve, I’ve already, I’m already trying to befriend the monster, you know?

Sarah: Oh, I know you are.

Taiyo: Oh, yeah. You know, Hey, Claude. You know, kind of, kind of love you, Claude. Just, just between you and me.

Sarah: [laughs] We’ll explore that in another episode. If you enjoyed this conversation, please subscribe to My Robot Teacher. Share the episode with a colleague, and please leave us a review on YouTube or Apple Podcasts. It really helps others find the show.

Taiyo: And before we go, we want to thank the California Education Learning Lab for supporting this podcast and for making conversations like this possible. Thank you.

Subscribe now

Leave a comment

Learning Lab is back on Substack!

In addition to guest columns, we’ll start sharing Q&A interviews, and project spotlights, as well as highlight interesting pieces (in case you missed them) that further our collective dialogue on California’s higher education ecosystem.

Here’s our first, “In case you missed it…”

• “Even in a marathon, the finish line is not the point”: Cal State LA’s Ji Y. Son explains why even in a marathon, the finish line is not the point – and what it means not to Waymo-ize education in the age of AI.

• Stanford’s Mitchell Stevens launches a “Learning Society” initiative designed to “transform the United States from the ‘schooled society’ of the 20th century, with its emphasis on formal educational credentials granted by schools and universities to a ‘learning society,’ where people can acquire the skills and knowledge they need in many different contexts and across the entire life course.” Stay tuned for a Q&A with Mitchell in April.

• Former Math for America President John Ewing opines on the “Crisis Mongering” that is now at UCSD’s doorstep regarding its admissions report detailing math preparation. Written in November 2025, Learning Lab recently became aware of the piece and is adding this to the annals of productive struggle with math education.

• Curious about insights gleaned from faculty who are leading AI Challenge projects? Check out these briefs: AI Literacy, AI Ethics & Responsibility, and AI Workforce Preparation, at Learning Lab’s AI Initiative webpage.

If you’ve published something that you want to share with the Learning Lab community, have a tip on a great resource or article, or want to submit a guest commentary for the Substack, please email info@calearninglab.org.

Guests:

• Pedro Morales-Almazán: Teaching Professor of Mathematics, UC Santa Cruz’s Physical & Biological Sciences Division; specializing in regularization problems arising in quantum field theory and asymptotic methods in number theory

• Julie Simons: Associate Professor of Teaching, UC Santa Cruz’s Baskin School of Engineering; specializing in applied mathematics with a focus on cellular motility, modeling, and computational simulation

Also available on: Apple / Spotify

COLD OPEN

Pedro Morales-Almazán: We don’t know what’s going to happen. And I think as academics we also should embrace that a little bit more.

Julie Simons: Once you add this huge probabilistic LLM, now you’re in a whole different regime where you have to be able to improvise.

CHAPTER 1 [00:18-6:40]

Sarah: Welcome back to My Robot Teacher. So Taiyo, when we last left off, you were waxing lyrical about Claude Code doing all the bureaucratic work that’s allowing you to now flip your classroom for the first time ever.

Taiyo: Yeah, the flipped classroom’s still going strong. Really incredible participation for my students. And maintaining my canvas shell now is SO easy because I have this Claude code generated script, which will grab links from my YouTube channel and automatically post all of the content to my canvas shell without me having to do anything. This is just a total game changer for me. Each of these announcements, if I’m at my best, would’ve taken me 30 minutes per class - 15 of which would be me feeling pity for myself as I have to wrangle all of these URLs and links and put them in a nicely formatted announcement. No longer, I can now offload all of this work to Claude Code has written for me using the YouTube API key and using the canvas API key. And it is magical.

Sarah: It SOUNDS magical. But what’s your classroom like now? Did you manifest the dream of the 2010s? Is everyone engaged in active learning in class now?

Taiyo: YES I DID! I did manifest the dream of the 2010s! Don’t mock me! I’ve got 25 students, they’re watching my lecture videos before class, taking a pre-class quiz to test their knowledge, and then coming to class and spending 50 minutes 3x a week - the entire time working, solving problems, discussing course content with their peers while I walk around and give a healthy mix of positive AND negative reinforcement.

Sarah: Oh you mean, when you shine a laser pointer at them and yell, “Where’s that number gonna go?”

Taiyo: Look, I’m not actually burning their retinas, okay?

Sarah: Oh that’s good to know you’re experimenting within Cal State approved boundaries.

Taiyo: But seriously, I really think this AI enriched flipped classroom idea is so powerful for this particular moment because you’re turning your classroom into a community of learners.

Sarah: Mmm, love that.

Taiyo: There’s a kind of social accountability that kicks in, something that doesn’t happen when all the cognition is happening alone with headphones on at midnight.

Sarah: Well or worse, like being outsourced to ChatGPT because you give them the problem set to take home and you have no guarantee they’re actually doing it themselves anymore.

Taiyo: This was a problem even before ChatGPT; the internet already made it way too easy to find ready-made solutions to your homework.

Sarah: Yeah

Taiyo: But now it’s even easier to give away your cognition entirely. And I think that’s a tragedy. And I feel like this intense active learning experience means they can’t completely offload their cognition because they’re being asked to demonstrate and own it publicly!

Sarah: Yeah, that also feels to me like accountability without like icky surveillance.

Taiyo: What you mean by that?

Sarah: It’s not, like, “I’m gonna put your problem set through an AI detector and hope that’s actually effective. It’s more like you are expected to come to class and demonstrate the cognitive moves that allow you to solve a problem and understand what you’re actually doing in this math class.

Taiyo: Oh yeah, for sure. I’ve got my students doing at the whiteboards, right, in these small groups, so they have to exercise judgment, um, constructive critique, analysis. They’re trying to understand what a peer who made a mistake was actually thinking, and then they have to demonstrate a kind of leadership in explaining where things went wrong. They’re problem-solving in teams, it’s like real teamwork. And I think it’s beautiful. These soft skills, social skills, whatever you want to call them, are going to be of paramount importance after college.

Sarah: Aww, you used to hate soft skills?

Taiyo: That is a malicious lie!

Sarah: [laughs] I think it’s interesting that we’re not even talking about AI in the classroom in a student-facing way. We’re talking about AI inasmuch as you were able to do this for the first time, it wasn’t such a leap with the amount of time you would have to spend setting up the infrastructure, but you’re not using AI in the classroom, right?

Taiyo: Yeah, I’m not using AI actively in the classroom, like in a student-facing way. I’m not closed off to it at all. But I’d still call what I’m doing right now “AI-enriched active learning.” But I think this is definitely different from what you’re doing in your classroom, which is extraordinarily student-facing AI use.

Sarah: Right, right. And in your case the class is AI-enriched and enables you to turn something that was passive learning into an active environment. I get what you’re saying. So how do you think it’s going to turn out?

Taiyo: I have no idea. We’ll see. I mean, I’ve got a basic underlying sense of what is effective. I know active learning is good. But really, when it comes to the day to day classroom dynamics now… I’m kinda just making it all up as I go along.

Sarah: Well, that’s fitting, because this episode is all about trying new things and adapting, and improvising inside the classroom with students (and sometimes with AI) in real time.

Taiyo: That’s right. We sat down with our friends Julie Simons and Pedro Morales from UC Santa Cruz to talk about what happens to the classroom when none of us quite knows what’s coming next. And maybe I should explain that I had a cold on the day I recorded this, which is why it sounds exactly like I have a cold in this episode. So anyway.

Sarah: Yeah, we didn’t train a very nasal AI bot on Taiyo’s corpus of work.

Taiyo: Did I really sound that bad?

Sarah: No.

Taiyo: Anyway, please enjoy this conversation with Julie and Pedro.

CHAPTER 2 [6:40 - 12:38]

Sarah: So we’re here right now with our former colleague, Julie Simons, who used to be our colleague down the hall at Cal Poly Maritime, and then defected to UC Santa Cruz.

Taiyo: Whoa, Sarah, relax.

Julie: Oops!

Sarah: And she’s here with one of her new work friends.

Taiyo: You know, Sarah Julie is allowed to make new friends. You’re, you realize that, right?

Sarah: I know, I know. But I do wanna know, Julie, what is the basis of this new work friendship?

Julie: Well, in addition to Pedro successfully recruiting me, he and I are both teaching professors at sort of companion math departments at UC Santa Cruz. I’m in the applied math department, and he is in the math department. So, we chat a lot about the issues our departments are facing. We co-teach some classes and things like that, so it’s been really great. The other thing that links us, I think, is our interest in DEI, you know, belonging and equity I think are really core to my values and thinking about how to reach all students, frankly.

Taiyo: So if you’re a teaching professor, does that mean you don’t have to do as much research? Or what does that mean in terms of like, in terms of what you do at your job?

Julie: Yeah, I mean, I think it just tilts the balance of all the things that I think every professor really has to do. We’re all involved in teaching scholarship or research and service at our institutions, and if you’re a teaching professor, you’re, you’re expected to lean in more on the teaching side, but we’re still doing research as well. Pedro and I both have active research projects, both within sort of our mathematical disciplines as well as thinking about research questions that relate to pedagogy.

Taiyo: Wow. Cool.

Pedro: I mean, on top of what Julie says about our roles as teaching professors, which actually I wanna highlight are pretty unique to the UC system, I think in other institutions that are getting more and more popular due to realizing how much this particular role is needed in higher education, um, because it’s not only about instruction, but also thinking about teaching and pedagogy in general, holistically. ‘cause we also have to think about our students’ career paths. Of course, we want to think about AI and the impact beyond just individual classes, but also in their own development as professionals. I would say that teaching professors, we are, we are very equipped to, to think about these things because it’s, it should be more than just checking to see if students are cheating or not to whatever that means. [00:13:48] And also to think how we can effectively use new technologies to promote their own learning.

Sarah: Oh, I love that. I totally agree. Julie, can you tell our audience a little bit about your stance on AI?

Julie: Yeah, I won’t say that I am. I think on, you know, one extreme or the other, like, I’m not an evangelist of AI. I also am very much not opposed, uh, when all this stuff came around, I remember Taiyo asking me if I had seen the whole, all of the news about ChatGPT and if I’d tried it out. And he was so excited. And his eyes got real big. And he is like, you’ve gotta try it. And I remember being like, okay, I’m gonna go try it because it’s Taiyo and what he tells me to do, usually I take seriously.

Taiyo: Aw.

Julie: So I checked it out, but I, I wasn’t like one of, I wasn’t like you and, um, like Taiyo and Sarah about this. Like both of you, I think, immediately got thinking “Like, oh, we could use it to do this and I wanna test these edge cases” and all of, all of these, uh, fun experiments that you have done. I wasn’t really on that page, but as I started playing with it, I started thinking about how students could interact with this, started thinking, like many people did, about what this might mean for personalization of education and tutoring. And more recently my feeling is also really centered around students and not letting students get left behind.

Sarah: Hmm.

Julie: I think a lot of our like STEM focused students tend to lean into technologies and get really excited about this, but they also might have, you know, the prerequisite skills to know how to test these tools and figure out how they work much more readily than some of our other students. Those other students are the ones that I spend a lot of time thinking about how to sort of support them and help in their personal development. And some of those students who are not exploring new tools have really great questions about ethics, about environmental stewardship and all of that. But what I worry about is whether those students get left behind by self-selecting out of it or by like society essentially selecting them out of it because of all sorts of barriers.

Sarah: And then they’re absent from the development of those technologies, yeah.

Julie: Exactly.

Sarah: Mmm. What also worries me is that self-selection could exacerbate achievement gaps, too: I see this happening now where some students opt out, others use it really effectively to support their learning, like, you know, doing stuff like using the quiz feature on ChatGPT or asking it to explain a concept three different ways (and confirming with me the one that clicks for them is actually accurate). And of course there are the risks to those who use it to avoid learning or offload it. Are you also seeing that reflected in the faculty conversation?

Julie: I think this is the debate that’s happening in, every campus right now; there’s faculty who are really promoting it, faculty who are really not, and then when I talk to people who are hiring our graduates afterwards, they’re all expecting students to be AI literate, not, and not just literate, but AI savvy. ] I’ll talk to people who are like, I expect my new hires to pass their ideas through five different LLMs, vet all of them, and then explain which one is the best. Now that’s a level of cognition and understanding that we used to expect, not from like entry level positions, but really after getting some experience. And so that’s now sort of leveling up what employers are gonna expect from new employees.

Sarah: Mmhmm.

CHAPTER 3 [12:39-15:27]

Taiyo: I think there’s always been a kind of debate in higher education about, you know, to what extent should we be in the service of the economy, of the workforce? Right? Uh, one of the roles certainly that we play is workforce development. Pedro, do you have any thoughts about that?

Pedro: I have a lot of thoughts about that.

Taiyo: Oh, please go for it.

Pedro: Maybe too many thoughts, but I think I have some practical thoughts about it. I think that we should acknowledge that a lot of our students come to higher education because of that. It’s a reality. But I also think that we have a responsibility, even more in the California system as public servants, to contribute to society and to support our students becoming better citizens. This means I don’t think it’s just about skill building; I think it is about holistically supporting them to become the best versions of themselves and, um, supporting their role in society. So I guess the question now translates into what kind of society are we going to have now with AI being an important actor - into the economy, into entertainment, into even politics. Part of what Julie I think is also addressing is how to develop that critical thinking that students will need in order to navigate this new world. I mean, it’s becoming a new world and one thing that I want to share, to say proudly, is that we don’t know what’s going to happen. And I think as academics we also should embrace that a little bit more. A lot of what we hear in our bubbles or not bubbles, is, highly speculative. We’re making things up and I think we are addressing our own beliefs, probably hopes and fears, but sometimes I believe as academics we should acknowledge that we don’t have the answers yet for what’s going to happen with AI in a lot of fields. And specifically to our conversation right now in learning. We don’t know to what extent this is going to impact student development, learning and behavior in general. So I, I do believe that it’s important for us to recognize again, that we are just still discovering and still figuring out how to better use AI in the different areas of humanity, but also within higher education.

CHAPTER 4 [15:28-24:12]

Sarah: One of the most compelling arguments I hear from colleagues who don’t want AI in their classrooms is that a work product is now so tantalizingly easy to generate that students will be unable to resist the temptation to reach for it, instead of building their own tolerance for persisting through not knowing. So how do we help students develop the tolerance to deal with uncertainty, and persist through not-knowing and also set expectations of what is and isn’t appropriate, while we’re also still figuring it out?

Pedro: So I personally would say it depends on the class, and what I try to doing in, in this is to usually, at the beginning of the term, to have a discussion with them, an honest discussion like, “Hey guys, what do you think would be an appropriate use of AI for this class?” And then it’s very interesting because you are not coming with a lot of assumptions, you’re actually generally interested in their, in their learning and this class was great because we have almost an entire lecture day spent on why would it be cheating, quote unquote, and what would be allowed?

Julie: I think that’s really great. Pedro. I think, you know, we’ve always tried to like incorporate students, or at least more recently, the trend has been incorporating students into understanding the course structure. A lot of things, you know, over the last maybe 15 years, faculty have moved towards like co-creating syllabi with students and, and talking through this. And so engaging them and understanding the learning process and our goals and allowing them to goal set in the class, I think is really helpful and, and really grounding for students to try to recenter them on the. The process of learning itself. Anyway, that was just bringing that up for me. But going back a little bit to Sarah’s question about, uh, struggle offloading, all of that struggle to LLMs and people’s concerns about this, I think, you know, a lot of us have been thinking about how to get our students to productively struggle for many years. A lot of our students, especially we see this in, in math classes, uh, where students will be like, “I was working on this problem for three hours last night.” And it, it’s, it’s a problem that should definitely not take that long. And the question then is: what are you calling working on the problem? That doesn’t sound like productive struggle to me, that doesn’t sound like you’re making any progress. And so getting students out of that productive struggle is actually one of the things that a lot of people who are trying to create AI tools specific to education right now are focused on how to get students unstuck. How to get them towards a struggle that actually leads somewhere instead of just feeling like demoralizing, a situation that doesn’t promote learning.

Sarah: Hmm, yeah, that’s something I’m definitely grappling with right now, you know, how do we create situations in the classroom that PROMOTE learning? It’s a perennial problem, I think. But what you said made me think of this social media post I saw this morning where someone said they have gone entirely back to pens and blue books because they, I think the quote was, they “don’t want to get AI slop.” And I think there’s an assumption there - well maybe more than one - like first of all, that all AI writing is slop. I think it takes some skill to get NON-sloppy outputs. But, more importantly for the point I’m trying to make here, I think it presupposes that thinking alone with that blank page in front of you is the most rigorous kind of learning. And just to clarify, I’m not anti-pen and paper at all. I actually do a bunch of hand-written assignments right now in my own class; but I tend to treat them kind of as one part of a scaffold, or one scaffold inside a larger writing process where students are allowed to use AI. But I’ve observed that when students use LLMs in a structured way, it can surface different cognitive “moves” in really interesting ways. But this is such a weird time where nobody has the answers yet, and so a lot of what I’m doing right now is intuition. It’s like patterns I’m noticing in my students kind of like hunches about what’s helping or not and a sense of what, what seems to be backfiring or workking. And so a lot of what I keep thinking about though is rather than thinking about like all AI use, good or bad, is really like to what extent does using an LLM help a student in a specific context and with what other scaffolding in place, right?

Pedro: Sometimes we as educators also have to acknowledge that we don’t know what’s the best way to use these tools. We are still adjusting to this. And sometimes I feel - I don’t know how you guys feel about this, but - I feel like we have this pressure to say one way or another, like whether AI is good for society or bad for society, or good for learning or bad for learning. And I personally like to replace that expression, that belief with to what extent? I think that a lot of the applications are a spectrum. They’re not binary. Just recognizing that, I think it’s, it’s something refreshing that we might need in academia to sometimes acknowledge that we don’t have the answers, that we don’t know certain things, and that should allow us to be more maybe humble towards discovery because in the end I do believe that academia has that responsibility. We are in pursuit of knowledge and discovery and curiosity. I think as educators we also should embrace that and should embrace that we are discovering, we are trying things and not everything is going to work out perfectly fine, but that’s okay. It’s part of the process.

Julie: You know, a lot of us academics think that academia should be this, this sort of bastion of curiosity and, and intellectual exploration and that we’re, we have this great culture of this, but one thing I’ve been thinking about, we’re all very discipline-specific. We’re trained in one department, but not just in one subject area. We’re all like so highly trained to be curious really in a narrow band and things outside of that band. Are still scary to us. We may think they’re interesting, like from afar, but to engage with that and admit just how little we know outside of where we’ve been trained to not know is, is a challenge. When I think about this translating to how we teach and how we interact with students and what topics we teach, I think there’s a lot of fear about not just change, but the loss of control of our classroom and of our student experience. Now, we’ve never really had control of the student experience outside of what, what little pieces we control. And of course, our students have always done different things outside of the classroom than what we expect. Thanks, Chegg! But, uh, but I think the idea for faculty right now who are maybe experiencing fear or resistance to this is, I don’t know what will happen. Not only do I not know whether it’s gonna, like how it’s gonna impact student learning, I also just don’t know how to react in that moment in class when students are engaging with these tools and they spit out something that is probabilistic in nature. So I can’t, there, this is not a deterministic process and that feels scary for some faculty who are used to saying, I’m gonna go today and teach this, this lesson plan that I’ve created. I know what questions I’m gonna ask students. I kind of know what questions they might have or where their mistakes might lie. Once you add this huge probabilistic LLM behind everything and allow sort of that, that loose, now you’re in a whole different regime where you have to be able to improvise. You have to be able to think on your feet and somehow address this. And that’s a scary, scary thought process I think for a lot of, a lot of folks.

Taiyo: Just real quick, one of the most pleasurable things that has come out of the ChatGPT moment is the five-year Chegg stock ticker to see what happened when ChatGPT was released. This was a stock that was trading at around a hundred dollars five years ago, and now it’s at 79 cents. So just saying…

Julie: Beautiful. [laughs]

CHAPTER 5 [24:13-35:06]

Sarah: This conversation about fear and loss of control, though, it’s making me think of that discussion that we had last summer, Julie, when you introduced us to Pedro, um, and we were talking about improv, and how you both like to think about it as a pedagogical tool or at least a faculty development, uh, thing. Um, I’d love to hear more about that.

Julie: Yeah, for sure. Well, so this, this all started, uh, with Pedro sharing with me that he’s been thinking a lot about improv and pedagogy, which was to me, like outside of my, my wheelhouse. I am not an improv person, not an actor or theater person. Those ideas scare me. But then I went to a conference last fall where we actually did an improv session and I was asked to lead one of those. I don’t know why, but I learned about all of these different improv activities that folks do, and it was amazing. It was like icebreakers on, on drugs or something, I don’t know, maybe on, on something really nice. But it was really transformative for the whole vibe of the conference. People got to know each other and it set a norm of community from the beginning. That was the most lovely conference I’ve ever been to. And if anybody listening has been to an academic conference, you know, it can be scary - a lot of people not talking to each other or only talking to those people they already know, and this totally flipped all of that upside down. But Pedro actually is the expert in the room. Um, he’s been doing improv for quite some time and has been leading an improv group in the math community at UC Santa Cruz.

Sarah: Pedro, do you see this as like a faculty development opportunity? Get more educators comfortable with improv and maybe people will be a little more adaptable and willing to experiment in the classroom?

Pedro: I see it more as a philosophy of life, honestly, because it is. And this is, this is to, to follow up what Julie was saying about that fear of losing control. It is a good antidote to that. If you’ve ever seen improv, you see how smooth things can be without any prior preparation, and I mean, preparation in a very loose way because the preparation has taken probably like years and years of skill building. It is a different type of preparation, and I honestly think that in our classrooms we now have to figure out a different kind of preparation. That’s the challenge that we have now with, with AI. The basic idea of improv, the way I like to think about it, is that it provides you with a framework and then you just basically focus on the little details on the go, which is doable and it actually allows for better collaborations. If I tell you the script of a sketch, that’s, that’s just me basically speaking through your voice. But if we both are improvising a scene, both of us are creating at the same time, we’re co-creating. It will be wonderful if we can have a co-created experience in the classroom where not only students, quote unquote are learning, but also the instructors - we are also learning, and I’ll be a little bit more radical because I actually, or maybe cynical is the best word, because I think that actually it is the opposite: It is usually the instructor that is learning and the students are just there observing the instructor learning. Again, maybe that’s a little bit radical, but I do think that both instructors and students can come to this commonplace to co-create learning.

Sarah: I guess it is a bit radical to talk about it that way, but it makes sense to me: we talk about “student learning” but often in the classroom what’s most visible is actually the instructor’s thinking. We’re the ones iterating in public, breaking ideas down, uh, adjusting explanations for different audiences. I agree we need to stop treating students like spectators of our expertise and instead treat them like apprentices in the types of cognitive habits that produce expertise. I also think, ​​I bet some people hear “improv” and assume that means no structure. But what you’re describing sounds more like a different kind of structure, or at least like it might require us to be more intentional about how we structure things and what we ask of students?

Pedro: I think that, um, at least my philosophy is that we can use this improv framework to be more successful at achieving our learning goals. But the, I guess the price to pay is that we have to be very clear: what do we actually want our students to accomplish? In other words, I think that we have to have a very clear idea of what it would mean for our students to learn something. And that’s the challenge that AI is bringing to our classrooms. I think that back in the day, um, we were all like, like The Karate Kid, like Mr. Miyagi, right? We were learning without knowing that we were learning something. as instructors, we were trying to teach them without telling them what was going on. Now we’re being confronted that we have to be very clear what are we doing to what purpose and how. So I would say, again, that’s my hopeful point of view of how AI is disrupting higher education is pushing us to really think and being very intentional about what is it our students will learn? How will they learn it, and how can we get there?

Sarah: Yeah, in the original Karate Kid, that “wax on, wax off” strategy assumes that students will trust your authority long enough to actually do the thing before they even get the point of it, which, I think, hasn’t been the case in higher ed for a while now, at least in my experience. So yeah, I think one of the ways we build that trust is to give a super clear articulation of what they’re learning. But that question - “How will they learn it, and how can we get there?” - that’s something, the exact route can’t be pre-determined for every student. So often in the classroom we teachers discover the “how” through iteration, and we might make missteps along the way, so it feels to me like part of our job now is also to teach students how to iterate through mistakes without shame.

Pedro: I think that’s also our role, right? Even more in math and in STEM. I do believe in the power of being wrong, and learning from that. I don’t think, I usually joke that math is the only subject in which having problems is a good thing. We want to have problems, and many times we want to fail from them. Failure, in my opinion, sadly, has a negative connotation where I do believe that failure is a sign of learning and progress, learning through failure.

Julie: Right. Pedro? This was when I brought my TAs that I was training, I was teaching a TA training class last fall, and we took a field trip to one of Pedro’s improv sessions. And, um, that was like the big point that I think was so lovely that you brought to our students was embracing failure and talking about how important that is in improv, but also in our classrooms, in our pedagogy and modeling that for our students. And as, as we were doing these improv activities, you know, you, you have different sort of, things that you’re asked to do, um, physically with your body and things you’re supposed to say and whatnot, and it goes quickly and you’re supposed to sort of maybe get in, I guess a flow state sort of with us and just not try to be in control. And it’s the thing that a lot of our students and faculty, frankly, it struggle with, I think in a lot of these realms because we’ve been taught that being correct is important. Knowing is important. And of course, like we, we want to know. We want that curiosity and we want to learn and develop. But letting go of shame about mistakes and embracing that and being able to laugh and sit and and celebrate the mistakes was something that was so beautiful about the improv session - like Pedro told us at one point, like, if you make a mistake, you have to like own it and like shout like “I made a mistake. Yay.” And move forward with that, you know? And we all got into this business of teaching. I think really focused on or motivated by facilitating growth in students. That’s what what we do this work for. It doesn’t pay us a lot. There’s a lot of other issues that come along with with choosing education as your profession, but it is so amazing to be able to facilitate student growth and getting them where they wanna be.

Taiyo: Do y’all ever make mistakes on purpose during your classes?

Sarah: Because Taiyo does.

Pedro: All the time!

Taiyo: Yeah, I do it too. And then what I like to do is try to gaslight my students into thinking that what I did wasn’t a mistake. They don’t like that very much.

Julie: They don’t like it.

Taiyo: No, they don’t like it when I do that. But I think it’s hilarious.

Julie: But that is also like, that’s again, that’s the kind of skillset that I want to be teaching them to do with the output of, of LLMs. Right?

Taiyo: Right.

Julie: Is to be critical, to be like, you’re spitting out this thing that Sure, it sounds right. You know, this sounds like something Taiyo would’ve told me in the classroom.

Taiyo: And be critical in particular of people in positions of authority. Like it’s incredibly important, right? Particularly these days. I mean, good god, just look at the news, right?

Sarah: Yeah. This takes me back to time and I always say: “AI Slop. What about human slop?”

Taiyo: Wait, are you calling? Wait, hold on. You’re not saying that I was doing human slop, were you, Sarah? You better not.

Sarah: No, you were doing intentional human slop, right? Oh, yeah. I love how you, I mean, I’ve like heard the legends of you doing this, of being like, “this is right, right? It’s right? Right?” And, and the students all sort of looking at each other.

Pedro: You know what, what really scares me is not artificial intelligence is human stupidity.

Sarah: Boom. Yep, yep.

Taiyo: There you go.

CHAPTER 6 [35:07-42:53]

Taiyo: You all are on kind of the front lines of math education, uh, in the 21st century. And one of the trends that I think we’ve all been feeling and observing in our students is an increasing lack of preparation, in our incoming students. For those that are not familiar, there was this report that came out of UCSD from their faculty senate, which reported on just this issue - that they’re seeing increasing deficits is the right word? And cracks in the foundational understanding of basic things, basic mathematics, basic writing skills, and that sort of thing. What do you think about that? Does AI have any role to play in helping educators with that issue? Particularly in light of some of the, uh, legislation, particularly in the state of California, which is making remediation a a sort of stigmatized, dirty word. What do you think about that?

Julie: Silence.

Taiyo: Oh, is this too spicy? Wait, hold on. Maybe this is too spicy?

Sarah: No, this is great. This is great. Literally like, oh, he’s going there.

Julie: Yep, yep. And we’re here.

Sarah: Maybe, maybe some context for listeners unfamiliar with, uh, the, the state of math education in the public system in California, a little background would be good.

Julie: Oh dear. Where did this start? Um, basically, gosh, what was it - 2017, Taiyo? The CSU was mandated to remove any remediation classes, and by that what we mean is classes that students were required to take that are not considered college level math, but that they were deemed necessary to take prior to getting to say a pre-calculus class or a calculus class - so what we consider college level class, that is credit bearing for students, meaning they actually get something on a transcript for it. Um, in the past with remediation, they had to take non-credit bearing classes sometimes, um, in order to finally get into credit bearing classes. And so that was essentially banned in the CSU around 2017. And then more recently, the state legislature mandated that in the community colleges. And so now basically across the entire state, this is essentially the deal, the UC system, because it typically has admitted students with more preparation, has not seen the same kind of level of impact, um, of those legislations. However, we are also seeing the same preparedness issues. I think every institution across the country is seeing math preparedness as a major, major problem. Our students are coming in, so the UC San Diego study that you cited, Taiyo, students are coming in not being able to do eighth grade math. How do you throw them in a calculus class without being able to do eighth grade math? You know, they don’t have those foundations and now we can’t do remediation. So a lot of the, the solution has been you add co-requisite courses somehow to support students through this, and there’s been some compelling evidence that that’s worked on some campuses and not on others. I think it really depends on how well it is supported, and it’s really hard to do this at scale.

Julie: I know on our campus at UC Santa Cruz, it’s a major challenge just figuring out how to place students appropriately and, and support for students. There’s a lot of different support mechanisms, but it’s hard to know like what is working and what is, what is not, and what the most effective practice is right now. So I do think like a lot of people are hoping maybe gen AI can solve that. But you can’t just assume that like, because it’s out there that this, you know, it’s like everybody thought 20 years ago that universities would cease to exist because you could learn everything for free on Coursera.

Sarah: Hmm.

Julie: That is not how most people learn. Honestly, I really do believe a lot of us learn best in community and in conversation and by relating to each other. And so, you know, one, one thing that we talk about is like, we still want that human connection even through AI. Like, you know, we want to be using AI and using different tools and preparing our students to be able to use these tools. But I also - like the soft skills are so important. What they need to be able to do is also be able to work with other humans. And that’s a challenge for some of our students right now.

Sarah: Wait, let me make sure I’m understanding this. So I always thought that the, like catch up route, if you, if you didn’t meet the expectations for, you know, high school math, was that you could go to community college, like get your pre-algebra foundation with really real like substantive support and then come back and take whatever courses required, um, like college algebra, calc, whatever. Is that still true? I thought our community college system was really incredible at that, um, aspect of education.

Julie: That certainly was like something that was an option for students. It used to be that… I’m really proud of our community college system in California. I think they do amazing work and I’m really proud of our transfer students that we get from community college. They are amazing.

Taiyo: Damn right!

Julie: And it used to be that I could tell students like, “Hey, actually, if you went and took this class at community college, you might get more support than at the CSU,” just because of how the systems work differently and what their aims are different. And so that now is no longer really the case because they’re, they’re subject to the same rules essentially. Um, they have different support mechanisms still, but it’s a real challenge. Like if you don’t know eighth grade math, there’s no class for you.

Sarah: Well, unless you can afford to pay private tutors or something. Wow.

Taiyo: Yeah. I know this is an often, cited use of AI or positive potential benefit of AI. And Julie, you, you mentioned how not everybody learns, you know, in front of a screen as you would have to be if you were working with an AI tutor. But do you think that there’s potential for AI tutors to sort of level the playing field? Because we know, like, and this was the case even when I was in high school 30 years ago, there’s like the, the SAT prep Industrial Complex, there’s the Kumon Industrial Complex that we know that rich families are able to put their kids, uh, into and, uh, and accelerate their learning through these kinds of extracurricular activities and such things were just not like that are, are just not an option for all families. Right, right. I, I wonder to what extent this, like personalized 24/7 infinitely patient, etc., etc. - you’ve heard the drill - AI tutor could level that kind of playing field. Do you have any thoughts about that?

Julie: I, I think there’s potentials with a lot of asterisks. I would say, a worry I have is of course, like even with AI, we have paid models, unpaid models, right? You get a very different product depending on what you are paying for or not paying for. So there’s an equity concern there. I mean, I think a lot of us who work in, um, the public education systems across the country and in California certainly are really hoping also to get first generation students and to help social mobility for, for the general public - that’s like one of, one of our key like sort of core values here. And so is that gonna, is that gonna reach that population? I’m not convinced. I think still there, there’s a lot of work to be done.

CHAPTER 7 [42:55-47:26]

Sarah: It brings me back to something Pedro said earlier about what is the purpose of higher education - that it’s framed sometimes as workforce preparation or, uh, also character development or holistic development of the whole self. And it feels to me that sometimes those things are positioned as mutually exclusive. I don’t think they are.

Pedro: Yeah, I would agree completely with that, Sarah. Um, I don’t think it’s, again, binary, it’s “to what extent,” right? It’s a spectrum. And what I think it’s more important for us as instructors and leaders in higher education is to be aware that these two things might, might not necessarily be competing against each other. We don’t have to choose either to do one or the other. One of the things that is happening right now in the industry is that a lot of entry level jobs are basically being outsourced to AI. So if you are left with a, again, economy that you only have, like middle to senior level jobs, how do you become a junior so you can progress? And I think that there needs to be this conversation between higher education and industry because it’s not effective for anyone. It is not effective for anyone - maybe in the short, in the short term is, is, is really good. It will boost, um, the economy or whatever. But then in the end, if all you’re left with are senior level jobs, how are students going to get those? So I do believe that there are a lot of things that we really need to be intentional about, again, going beyond the classroom and going beyond tools to, uh, improve productivity, but also what is the role of AI in society in some sense, and how we can be intentional about that.

Julie: Yeah, I think like returning to the equity issue and thinking one thing that keeps me up at night, um, you know, faculty love to worry about all sorts of things affecting our students, but I meet a lot of students who are anti-AI or just not convinced that this is the direction the world should be going and what I worry about is losing, like this is different than other technologies. AI is being trained on its users, and so if the people who are opting out of using gen AI are the ones who are most ethically minded, who care most about society, our planet. Then these tools are going to miss out on those voices. And those are the voices I think that are the most critical right now in leading us where we need to go. This presents a challenge, right? You have a student who doesn’t wanna use gen AI because they think it’s an unethical tool to use, it’s destroying, destroying our planet environmentally, data privacy issues, and ultimately, potentially the thought is it might destroy humanity, right? Or overtake humanity. And so do we wanna contribute to that? But those are the people that like, they, we need them in these conversations. We need them to know about how these tools work. And choosing not to engage with the tools means missing out on understanding the tools too.

Taiyo: I definitely agree with that. It is really important to have a diversity of, uh, perspectives and voices, embodied in these AI systems that are currently being developed. I think, though, an anti-AI stance in higher education, an anti-AI stance, that is not necessarily an ethical stance in my opinion. I mean, it’s ethical in the sense that it is a sort of normative claim, but to me, to deny your students access to this world-shaking technology that’s going to strengthen them. Forget about workplace stuff. Forget about the economy. Just strengthen your ability to learn things. And that’s been my experience with AI. I think that to deny them wholesale from that technology, I don’t consider that to be an ethical move. People might believe that, uh, but I would want to have that argument. And I think it’s a really critical, important argument to have, and I don’t think that an outright ban on AI is responsive to the fact that that is an argument, there is an argument to be had there. Anyway…

Julie: [laughs]

CHAPTER 8 [47:27-57:40]

Taiyo: Oh my God. Why was that so awkward? I think I made it awkward, honestly, right, Sarah?

Sarah: Yeah, I think so. Um, because you basically just called every educator who has a no AI policy on their syllabus unethical.

Taiyo: Wait, wait a minute. Did I really come down that hard?

Sarah: I don’t know. I mean, I guess we’ll find out by the amount of hate mail we get.

Taiyo: Oh well. Bring it on. I’m ready. I respond really well to negative stimulation, so…

Sarah: We’ve established that.

Taiyo: Looking forward to it.

Sarah: Well, I mean, I think I agree at least with part of what you were saying, that there is an argument to be had, and that a lot of the blanket “AI-has-no-place-in-higher-ed” arguments are based on an assumption that using AI can only outsource cognitive labor and, and not help not enhance it.

Taiyo: Yeah.

Sarah: Or not support students, right?

Taiyo: Well, yeah, and I get that and I also understand Julie’s point of view where, you know, she was talking about how when you introduce this sort of probabilistic, you know, entity into your classroom, it can be difficult because there’s a feeling in which some degree of control which you may have had in your classroom will be lost.

Sarah: Mm-hmm. Yeah. That part was really interesting to me. You know, I was thinking a lot about this, the loss of control, because I think I pride myself on being very flexible and adaptable in the classroom. And yet for some reason, even though it’s been about a month since you first told me about Claude Code and I was really excited and I was like, show me how to do it - something about, I don’t know why I’ve been resistant, I haven’t done it. And I know it’s gonna save me so much time every day in class. I have a really great talkative section in my critical thinking class this semester. And for some reason, I can’t pry myself away from the model of where I do like a 10 minute lecture at the beginning of every class period. This is a class I’ve taught a million times. It’s like one of my favorite classes to teach. I could make video lectures so easily for this class.

Taiyo: Yeah.

Sarah: And I don’t know why I haven’t taken that leap. And something about what Julie said about the control thing makes me. I don’t know. I think that’s sort of at the root of it in some way.

Taiyo: Yeah, I think that’s there for me too, in addition to the logistical nightmare that is flipping the classroom. I think the other thing that really deterred me from it was that loss of control. Um, as a departure from such a familiar model for myself of how teaching and learning should happen.

Sarah: Yeah.

Taiyo: When I was a student, I really admired so much my professors who were able to, you know, weave together ideas and such, you know, beautiful tapestries of knowledge. And I thought that that was the way everybody learned and had those incredible aha moments and the ecstasy of epiphany and all of that kind of stuff, right?

Sarah: Right.

Taiyo: But that’s just not the case, right? And science is like, the science of education has really established that active learning is really what works best for the majority of our students. And it’s, I think, really important we all take heed of that.

Sarah: Yeah. One thing I keep coming back to is like right now, you know, reading all this stuff, the really promising research coming out about how impactful AI tutors can be as a supplement to a course.

Taiyo: Mm-hmm.

Sarah: One of the things I think because I’m experimenting as I go, I really like to have the student engagement with AI happen in the classroom. I just started this semester, giving them guided prompts to do is like quote unquote homework. But then we go over the outputs in class. But so far, for the past few years, most of the workshops I’ve done where they’re doing AI generated writing or AI assisted writing, they’re doing it in class and that is, it’s totally active learning. I wish I had more time for that.

Taiyo: Mm-hmm.

Sarah: I think I have to do it. Taiyo, I think I actually have to follow through and now I’m saying it on the air and so I don’t wanna be embarrassed in a month’s time.

Taiyo: Well, let’s, let’s do this together. I, I like,

Sarah: Oh my God.

Taiyo: Let’s figure out ways to, and, and share with one another about ways that we can make this flipped classroom thing - the AI enriched flipped classroom - let’s talk about ways that we can make that work. What do you think?

Sarah: I think that would be an awesome idea. I mean, I need a motivator, right? I need a friend to do this with me.

Taiyo: Well, hello right here!

Sarah: [laughs] I know you’re already doing it. Well, I mean, here’s an interesting thing. Are you doing anything with AI tutoring with that class?

Taiyo: I’m not.

Sarah: Because this I think would be a challenge for both of us. Like right now, I see you doing it. I’d call it like an instructor-centered use of ai. Yes. Where you’re using it to like make your life easier so that you can do something. That was available to you pre AI that you just couldn’t do because of like logistics and time, right?

Taiyo: That’s right.

Sarah: Whereas I, for whatever reason, even though I know how easy it would be, have not done that. But I feel like I have been doing pretty cool stuff in the classroom with LLMs and I wish I had more time to help guide students through the process of engaging with LLMs. And what would free that up is if I had the lecture stuff. Because here’s the, the kicker that I’m realizing right now. The thing that would really I think take stuff to the next level in my class is that if they were able to like watch the 10 minute lecture and then have a conversation with an LLM, maybe a custom trained ChatGPT on the content on my course. But if they are able to have like that pre-class kind of dialogue, low stakes with the LLM and then class time is the, you know, like, it feels like a much more like honed springboard. That’s a weird metaphor, but a much better springboard.

Something where they don’t, they don’t come to class like currently - and how I’ve always done it, and again, I think I’ve done, you know, this class is, is like a favorite of mine and my students, if evals are to be trusted, but I see them kind of overwhelmed after my 10 minute lecture, right. I think for a long time I’ve been like, “yes, that confusion in your face is the…” right, like I’ve been doing the same thing of treating the like struggle as a kind of moral high ground as opposed to experimenting with what it would look like to have them have an interaction with an LLM where they can say like, what the hell is she talking about when she says this, right?

Taiyo: You give them space and time.

Sarah: Yeah.

Taiyo: And, uh, a sort of a safe interlocutor. Mm-hmm. A low stakes discussion with an LLM to rehearse their ideas before they come to class.

Sarah: Right.

Taiyo: And share publicly

Sarah: Right.

Taiyo: Their ideas with the rest of the class.

Sarah: Yeah, exactly.

Taiyo: That makes a lot of sense to me.

Sarah: I’ve done that to an extent, like in critical thinking right now, where we’re, we’re working on, you know, critical reading and explicating, textual details of stuff that they’re reading. Everything from poems to advertisements. And one of the things I started doing that has made discussion much more active than I think it’s ever been is having them ask an LLM, what are the implications or connotations of this word? What could they be? And I ask them to try it themselves first, but then to write down, ask for 10 and pick the three that are the most compelling, and then come to class being prepared to explain why they think it’s most compelling. And the reason that I think, you know, you could say that’s offloading the work of them sitting there struggling and thinking about what the connotations are. What that ignores is that like realistically, probably more than half the class for the whole time I’ve taught this comes to class with like, I don’t know what the connotations are, right?

Taiyo: Mm-hmm.

Sarah: And we work through it together and that’s been really awesome. And I love seeing that where they really struggle and then like over time through class discussion, they get it, but it usually takes weeks. And I’m wondering if it could be accelerated in a way that did not offload something but enriched something by giving them more examples, more models, and then also the, that we’d be using class time to like pick apart the outputs and build on it.

Taiyo: That’s, I mean, that sounds incredible to me, honestly.

Sarah: All right, let’s do it.

Taiyo: It really does. It really sounds great.

Sarah: All right.

Taiyo: You’re committed then.

Sarah: I’m committed.

Taiyo: Because this is going on the podcast. It’s going to be, this podcast is gonna go out to our thousands of listeners. You are really committing publicly to this, right Sarah?

Sarah: I guess so. You know, I love the pressure of a deadline and social anxiety.

Taiyo: Well, we’re gonna document all of this, every step along the way on this podcast!

Sarah: This is like the story of our lives: signing up for more work. We’ll let you know how it goes.

Taiyo: Well, maybe. Okay. Honestly, I was gonna say no, no more work because of Claude Code, but I mean, this is gonna take some thinking in some planning. It’s gonna be awesome.

Sarah: All right, well, we’ll see what happens. Until next time, I’m Sarah Senk.

Taiyo: and I’m Taiyo Inoue.

Sarah: This has been My Robot Teacher, brought to you by the California Education Learning Lab. If you haven’t already, please subscribe to our YouTube Channel!

Taiyo: Leave a review on Apple, Apple Podcast. Please?!

Sarah: Leave a review on “Ah-pull?”

Taiyo: I’m sorry.

Sarah: Please leave us a review.

Taiyo: On Apple Podcasts.

Sarah: Especially if you were that guy at that conference we went to who said this was better than Hard Fork.

Taiyo: Oh yeah, please that guy. Shout out to that guy! [laughter]

Below is the full transcript of Episode 9 of My Robot Teacher (lightly edited for clarity and concision).

Guest:

• Jon Sack: Associate Professor, Department of Physiology and Membrane Biology, UC Davis School of Medicine

Also available on: Apple / Spotify

INTRODUCTION

CHAPTER 1 (00:00-6:27)

Taiyo: Welcome back to My Robot Teacher.

Sarah: And welcome back educators from Winter Break. Taiyo, what did you do?

Taiyo: Oh my God, Sarah. So no joke. I had Claude Code create my entire canvas page for my differential equations course. I’m not even kidding.

Sarah: Wait. Design it or create it?

Taiyo: Well, you know, update it, maintain it, structure it. See, I gave Claude Code the academic calendar for spring 2026. I gave it my syllabus with the topics that I wanted to cover. And I gave it a link to the open educational resource online textbook that I’m using for the course. And I told Claude Code, please construct a day -by -day schedule for my course and build the entire canvas shelf for it. And it did it. It mapped out every single day of the semester. It created pre-class reading quizzes for students to check their knowledge. It organized everything into modules. Put the quizzes underneath the correct modules. It did the whole thing.

Sarah: Okay, so it didn’t like make the course for you. It took your materials, organized the course. I mean, this is incredible for you because of how much you hate organizing.

Taiyo: Yeah, you know, I’m the one that designed the course, right? I am, after all, the instructor of the course. But what I hate figuring out is how I’m going to take all of those topics and map them onto specific days. And you know, students love that kind of structure. They love knowing that on this or that day, we’re going to talk about that or this topic. You know what I mean?

Sarah: Totally, totally. Yeah. But you’re not ready to set up the scaffolding in January for May for the class.

Taiyo: I lack the mastery over time and space to be able to do this sort of thing effectively. I mean, you know how I am with time, right, Sarah? I’m so bad with dates and I don’t have a solid understanding of the difference between past, present, and future. I have real problems around all of that.

Sarah: You also hate the drudgery work of sitting there and hitting “edit module, add page,” and actually like putting all of that shit into your campus page if you’re, you’re teaching it for the first time, right

Taiyo: Oh my God, I hate it so much. And it’s one of the reasons why I haven’t done the heavy lift of flipping my classroom, which, by the way, I’m doing now for the very first time in my class, or the very first time in my career, in my teaching. Because, you know, I’m totally for flipping my courses, but the amount of structure that’s required for students to feel confident that, you know, that the course is going to be good for them is just immense. And that requires so much pre -planning, so much execution, so many clickings of buttons and managing of dates. Oh, my God. But Claude Code is doing all of that for me, Sarah. It’s doing it for me.

Sarah: Did you splurge for the, like super expense, the, the top of the line one for Claude code?

Taiyo: Not yet. I’m just using the $20 a month one. But you know what, Sarah? I’m pretty sure we’re going to have to like, we’re going to have to go in on that $200 a

month Claude account. You know that, right?

Sarah: I can’t believe if your $20 a month one was able to like make the entire course for you and there were no hallucinations, nothing?

Taiyo: Well, nothing I’ve detected yet and I’m going through meticulously as one does with a fine tooth comb as things are coming out. But so far, it’s been beautiful and completely error-free.

Sarah: If that’s what the $20 a month one… I would be very curious to trial it.

Taiyo: Yeah, you know, I really think we should. I think we should split an account and, you know, $100 a month apiece. What do you think?

Sarah: Let’s ask Anthropic first. I’m shameless. Let’s just, let’s just be like, “Hey guys, we would love to demo this.”

Taiyo: That’s right, we would love to, we would love to demo this, but we are CSU faculty. We don’t have the, the, uh, near infinite coffers that you all do. Could you give us a little taste? Please.

Sarah: [laughs] All right. We are gonna have to talk more about this later for like immediate, you know, hot tips for social media accounts because this, there’s so much here.

Taiyo: Oh yeah, for sure. And listen, audience, if you haven’t already, please subscribe to our YouTube channel because we’re gonna be putting out quite a few videos about practical tips like this.

Sarah: Oh yeah. Uh, see, now that we’re full of energy at the start of the spring 2026 semester, we’re gonna start posting short clips about what we’re doing right now in our classrooms, what we’re testing in the wild in this, you know, experiment of higher education in 2026. But we are here today to talk about a discussion that we recorded. In November with UC Davis Biophysicist, Jon Sack, who is an associate professor and researcher in the Department of Physiology and Membrane Biology at the UC Davis School of Medicine. Full disclosure, he’s also a friend of Taiyo’s.

Taiyo: Yeah, that’s right. That’s right. Yeah. Our kids went to the same preschool, known him for over 10 years now. We just met at some point, uh, uh, coincidentally and had some informal conversations about AI as one does right,, and I just thought he would be a really great guest because he is a scientist and a science educator and he’s got really interesting things to say about how AI’s gonna impact both the business of science, but also of science education.

Sarah: And because Jon is a scientist and you know, so far on My Robot Teacher, we’ve talked to a lot of humanists and social scientists and data scientists most recently. And so we were interested in particular in what Jon thought AI was doing in science education. We found a lot of common ground because I think across any classroom right now, the worry is really obvious that if students can outsource the work and, and still turn in something that looks very plausibly right, what do we do about that? And more importantly, how do we teach them to recognize what is right if everything kind of sounds good.

Taiyo: Thank you to our sponsor, the California Education Learning Lab for sponsoring this episode.

Sarah: One of the best things about working with them is the ability to connect with faculty across all three segments of California public higher education, and to think in interdisciplinary ways about how we are bringing technology to bear in the classroom.

Taiyo: Now onto the episode. Please enjoy.

PART 1: SCIENCE EDUCATION

CHAPTER 2 (6:28-10:14)

Taiyo: So Jon, you know, it seems like everybody has their own story of what their ChatGPT moment was. What was that like for you?

Jon: The initial reaction was like, wow, this can really deal with all the dumb crap in a way that I can communicate with, you know I think the initial experience I had was that it’s, you know, it’s as, uh, it’s as effective as an enthusiastic, you know, kind of naive trainee, or, you know, assistant.

Sarah: Who really wants to please you.

Jon: Yeah, exactly. Even more and more so now, right. And yeah, but the, that it could, that it could offload so many of the, of the, you know, time consuming, you know, mind numbing tasks, you know, quite readily. And, but, but also just the, just interacting with it and seeing what it knows what it could understand. It could not, you know, it couldn’t do a regression analysis to save its life at first, but now it can do that quite well. I was also floored. I came in later into the game ‘cause I read about everyone else having these religious experiences upon, you know, talking to ChatGPT. But I immediately wanted to start using it for everything possible just to see what it could be used for. I suggested to, you know, all of my students that they use it for, try it for. Everything and see how they can augment their capabilities with it, which has had, you know, really a, a, a wide range of unexpected manifestations, you know, from lots of disappointment to solutions in the lab being made, you know, much more accurately without mistakes. Sometimes students expressing frustration with me as an advisor because I’m not as positive and supportive as the ChatGPT.

Taiyo: Wait, what? Really? Because I know you to be a very positive and supportive guy.

Jon: Uh, thank you. Yeah, I try. Yeah. I believe, and, and I think, you know, in the, you know, it, that’s a wonderful thing that I think it, you know, it, it adapted, which is that it, it turns towards you as much as possible. And the, uh, the large language models I’ve worked with, you know, the kind of, you know, consumer facing ones, they do a very good job of turning towards - you know, even when they’re saying no, it’s a very gentle “no,” “it’s so true,” “Great question!” And “no, there doesn’t seem to be a correlation there,” but when there, when you ask a question where there is a correlation, it rewards you with turning towards you very positively. And I actively, you know, try to do this ‘cause people like it, but ChatGPTs is better at.

Sarah: It’s full of patience, it’s never hangry, it’s never tired.

Jon: No, none of this. Uh, and yeah, and when a trainee comes into a good idea. That the, the AI will agree with and, and, and reinforce that it’s a good idea. And I find myself being in the strange place of not just doing the normal thing of saying, “Yeah, you know, I told you that. I think, you know, please come to me with ideas, you know. A small fraction of them will probably be good ideas, but we should talk about them,” There’s this extra level where, where people are being convinced their ideas are good ideas. And I as an advisor, wind up naysay or casting shade or suggesting. They do some more, you know, reality testing of these ideas that, that students are already convinced of are good ideas due to those interactions. So it creates, uh, frustrations in unexpected ways.

CHAPTER 3 (10:15-12:44)

Taiyo: Interesting. Yeah. You know, Sarah and I, one of the big themes in this podcast is in thinking about what qualities or what kinds of, um, maybe you want to even call them virtues that we wanna see in our students to be able to cope with this new world where we have AI systems proliferating and becoming more and more common and impacting our lives in highly non-trivial ways. Like how can we make them resilient to the kind of maybe sycophancy that I’m hearing you describe coming from these LLMs - that they’re getting gassed up by, you know, all the kind positivity that might not have any real basis in reality. And it, you know, it takes somebody like yourself who’s a deep expert in, in these matters to. Have to do the kind of annoying work or frustrating work of throwing cold water on students’ dreams and that sort of thing.

Jon: [laughs] Exactly. You want, you, you want, you want ideas that may not go anywhere to, to fail fast, as they say.

Taiyo: Absolutely.

Jon: And they can kind of propagate for longer. Yeah, and I, I think core resiliency, I think is really the number one trait that I’ve noticed that at least in the sciences that we’ll get. You know, that will help you continue because that’s what science is all about It’s about hard reality testing of your favorite ideas and watching nothing emerge from your experiment, uh, you know, again and again and again, and being resilient enough to go again and again to the, the gallows of reality testing, and then, so that you’re, you’re still present and functional when you almost stumble upon something that really, really does work or sink in, or you actually have figured out, uh, how a process works in the, in the body, in our, in our physiology. You, you have that aha moment where you. Really have a better idea of how say, you know, an ion channel we study, integrates its information. You know why a specific drug that’s very powerful, what is the core secret that makes it work? And you want to be able to put out a lot of ideas and relentlessly, you know, select among them, you know, again and again and again to see all your favorite hypotheses squashed. So yeah, resilience.

CHAPTER 4 (12:45 - 16:08)

Sarah: So I have a two part question here. One is: do you agree that [resilience is] a fundamental part of scientific literacy, let’s say. What other elements are part of [scientific literacy] that I’m missing? And then the second part is: How have those things changed or how has pressure been put on those skills since ChatGPT - since students can now have it write their lab reports for them, what kind of work are they outsourcing and what are the risks you see in terms of how it might be damaging people’s scientific literacy?

Jon: Those are good questions. The first one - the fundamental question of having a scientific hypothesis: A scientific method is not to prove your hypothesis, it’s to try as hard as you can disprove your hypothesis - that I think is what you’re alluding to - to find every way you can of seeing if it’s wrong. And the hard part is to rejoice when you’ve killed it - when you’ve proven it wrong. In writing sometimes, in literature, there’s the expression that you must “kill your darlings.”

Sarah: Yes. That’s exactly what I was thinking.

Jon: I teach that you know, as you know, as part of the scientific method is that your most brilliant idea - you want to, you know, you want to shut them down as fast as possible. And if you’re unable to do it, if you’re unable to shut them down, then that’s a success. And I think the way that, at least with ChatGPT you can train it to do that, you know, to help you with that - to give you, you know, positive feedback for destroying the ideas that you hold most dear. Because if you destroy it, that’s a success. And it’s almost like if you haven’t obliterated your hypothesis, then it’s kind of a mini failure. But you should say, you know, that’s good. You know, keep on trying. I love that. The worst thing it could do is to say “Success. You failed to disprove your hypothesis. You’re a winner. You’re done. You know, you should write your paper now.” You want to, you want to keep chipping away at it. And the good thing is that it will, you know, reinforce what you want. It wants to please you, so you need to train it. As a scientist, that’s not the typical way people wanna be treated.

Sarah: Right. Oh, that’s amazing though. I feel like that’s a really great practical tip - that for context engineering or prompt engineering, like if you’re a student in a science class he first thing you say in your conversation with ChatGPT is “Your number one goal here is to give me praise when I obliterate the things that I clearly am invested in personally.”

Jon: [laughs] It can work with that.

Taiyo: I mean, this is a really important thing for me. This is why I’ve told you in the past that I respond really well to negative reinforcement.

Sarah: Yeah.

Taiyo: I really do. Like I respond much better to negative reinforcement than I do to positive reinforcement.

Jon: That’s terrible.

Sarah: [laughs]

Taiyo: I know, I know. Whatever. It’s fine.

Jon: You’re a bad person.

Taiyo: [laughs] Thanks. Thank you very much, Jonathan. But it gets to the points now…

Sarah: This is like when I yell at him like this in the hallway and people are like, what a psychopath. But I’m like, “He likes it!”

Taiyo: [laughs]

Jon: [laughs] I’m starting to wonder why I’ve ever liked you.

Taiyo: Yeah, well, you know,

Sarah: You’re just motivating him.

Taiyo: That’s very interesting - the feelings that I’m feeling right now. Anyway…

CHAPTER 5 (16:09-19:26)

Jon: Yeah. I guess the question, yeah, about doing the hard work.

Sarah: Mm-hmm.

Jon: And I guess what I see is that if the hard work is, you know, kind of a, a side effect of some task then, and you can offload the hard work to get the task done, you’ve still got the task done. And whereas if the task itself is doing the hard work, then - if that’s actually the point of achieving it - then if you offload it, then you’re not, you know, kind of doing the hard work.

Sarah: Yeah. I love that distinction.

Jon: So it’s kind of, it’s the, the framing there, that’s different.

Sarah: Right, so can you think of an example of in, um, like a, say a college level intro to biophysics class, maybe where there is something that students perceive as busy work - work they shouldn’t have to be doing. And but something that you think as the expert in the field is like fundamental to their understanding of this, that maybe the question is just reframe and explain. The point is the process. The point is to feel, this is hard.

Jon: Uh, to feel that it’s hard. I don’t know that it’s to feel that it’s hard. It’s to… So what I think of is math. Essentially I teach physiologists. I teach first year molecular, cellular and integrated physiology students who come into come into UC Davis, and what I teach them is hard thermodynamics, energetics, ligand receptor binding. And the way I get them to engage with it, is to have them work through math, conceptually. You know, the addition, subtraction - that’s not the point. What mathematics is, in my mind, is it’s logical relations between things. Right? And to get them immersed in those logical relations and see how they work. And I tell them to take the perspective of a molecule. I think that is the hard work - to imagine your inputs coming in and how that shapes the decision you can make - what the concentration of a drug or a neurotransmitter is, and based on what that is, what kind of decision are you gonna make? Are you gonna bind it? Are you gonna, not, how much of the time are you gonna bind it? Are you gonna not? And all of my questions you can feed into ChatGPT, and get the right answer out, but you don’t actually understand, you know, anything. All you understand is that you can get the right answer out. So I try to talk with them and to have them adopt that perspective. And they, and I’ve had them do a lot of math as the precursor to kind of getting there. What they don’t know is that on their midterm, they’re not gonna do any math. They’re merely gonna be explaining the relationships between things, between inputs and and outputs, and whether they can kind of verbalize that.

Sarah: Mm-hmm.

CHAPTER 6 (19:27-25:25)

Jon: In this particular course, the consensus had been to have testing be, you know, in a room without phones - like pens and paper - so that they are left with the devices, you know, with their, their thought processes. It has been an interesting, integrated mix: I encourage them to use ChatGPT or to use large language models to help understand things while at the same time the course forbids them from using these same large language models to answer questions.

Sarah: So these classes are intended to train people in the basics of scientific work, right? I mean, in my field there’s a lot of people commenting on how students need to learn the foundations before they mess around with LLMs. I’m wondering if that’s similar in the sciences.

Jon: We are - as scientists - one of the very helpful principles or something that’s generally embraced and shared is that if you’re a scientist studying a process, you should and really must look at all the new technology that’s available and bring it to bear on the problem because you get rewarded for how much new knowledge you unearth. And if you, when you have new technologies, new ways of seeing, then you’re gonna see things before anybody else does and understand them before anybody else does. And that’s what you get rewarded for in the sciences. And at the same time, many scientists are… they’re, you know, resistant to this wonderful new, available technology of, uh, large language models. And I think, you know, relatively quickly they’re adapting and adopting. It’s taking some time on the timescale of years, surprisingly. But I think what we need to teach our students is to bring every tool they have to bear onto their problems. Because like the large language models as we know it, these will be obsolete, you know, probably, you know, on the very short scale. And so teaching them how to use, you know, ChatGPT 5, I don’t see a great value in that, but encouraging them to go out and use whatever is available and to harness that technological power and to wrangle it in such a fashion that it can be useful to them - to help make a new, fundamentally new tool work for them. I think that’s valuable and that’s what we can teach them.

Sarah: Yeah, for sure. You know, we all know that LLMs can generate really convincing but wrong ideas, which is a problem for people totally new to a field who don’t yet have, maybe, the expertise or competence to discern among plausible sounding solutions, ideas, whatever. So as a student today, how do you learn to pick the good hypothesis from all the bad ones, especially if like so many of our students today, you’re fixated on getting the right answer?

Jon: If you’re in a field where you can get the right hypothesis every time, then you’re not really pushing yourself hard enough. The cutting edge of science is always gonna be in an area where there’s lots of ideas, but we don’t know what the right one is. And how to select the good, find the good idea, or the quote unquote right idea amongst all the bad ideas is fundamental to what we do as scientists and can best help the next generation by, you know, training them to do this. But the world in the, the kind of structure, of the topology of the world in which we do this now is changing rapidly with all this computational power. And one of the great things I think about the large language models is that they hallucinate - they throw out things that seem really right, really good ideas. And we need to be able to identify those hallucinations and, and reject them and to have the mentation capabilities to figure out how to do that. And the playing field will keep shifting and changing dramatically. So I think AI - these large language models are extremely helpful and they have unexpected consequences - as everything does, that’s new that that, that you work with. And I think for students, again, resilience is important. I really think also, for everybody generally, but especially for students, is to be aware of the limitations of what new technologies they’re, that they’re they’re working with. And I think it’s important for scientists especially to go boldly into uncharted territory and to be continuously taking stock of where it fails. And what it can’t do for them. And to be aware of those dangers. ‘cause in every, with every technological revolution that I’ve seen in, in science, there’s great new technology that emerges and there’s, you know, a lot of really profound findings using it - some of which later turn out to be wrong because there was things you didn’t understand about the new technology that you didn’t control for that led you to, led you astray. And to be ready to harness all the capabilities of every new technology that comes out, and to not be shy about addressing its limitations.

PART 2: AI IN RESEARCH

CHAPTER 7 (25:26-28:21)

Sarah: So, Jon, we’d love to talk to you about the implications of AI for research in your fields. You’re a biophysicist. Tell us more about what you study, where you are.

Jon: So yeah, so I’m, I’m an associate professor at the University of California Davis in the departments of Physiology and membrane biology. I’m in the medical school. I’m a basic researcher. I study the way molecules in our electrical system, in our body - primarily in our, in our neurons, the way molecules make stochastic decisions, the way they integrate inputs and turn those inputs into an electrical signal. And these electrical signals are important really everywhere for secretion of hormones, for the beat of the heart, for the contraction of muscles, and to create the electrical signals that propagate through our neurons and nerves and, and brains. The electrical work is readily studied in a lot of ways because it’s electricity and we can study it very, very well. And it’s interesting to me because it’s what makes our nervous system run; it’s how we generate and propagate electrical signals from one part of our body to another, from one part of a neuron to another, and how a neuron, a cell, how the molecules in that cell make decisions about whether to make the voltage more positive or more negative in a cell. And those types of decisions from many, many types of proteins and parts of our cells and bodies - eventually we build off of these little bricks into the form that is a cell or a neuron. And when you get neurons working together, you get a, a neural network and a, and a system that creates us essentially.

Sarah: What’s really interesting to me about that is the idea of something that is like a cell in your body - so not something that we think of as conscious - making decisions. So this sounds a little bit like the way we talk about LLMs.

Jon: Yeah, they’re, they’re, they’re highly analogous and LLMs, the neural network underlying network, uh, LLMs is inspired by the way neurons in our, in our brains and nervous systems work, which is they integrate inputs and with a, a stochastic, you know, weighting, they make a decision about what to output. And I think that underlying architecture has some incredibly powerful informational aspects, and so they’re related, but we, I don’t, I don’t study computational neural networks. I study at a very basic level the way inputs with stochastic weights are combined to create an output, which is analogous in many ways to the way a neural network works.

Sarah: Hmm.

CHAPTER 8 (28:22-30:36) - TERMINOLOGY BREAK

Sarah: Quick break to go over some terminology, which we always like to do for people new to LLMs. So, Taiyo, “stochastic” basically means probability-driven, right?

Taiyo: Yeah, exactly. exactly. And when you get to understanding reality at smaller and smaller scales, what you learn is that the role that probability and chance plays - well, it becomes baked into the fabric of reality. So, like, maybe you’ve heard of quantum mechanics and quantum physics and that sort of thing. And there, things are are inherently unpredictable and you can no longer just know that something is going to happen, you rather can only know that there’s going to be a distribution of possibilities of things that can happen. And that’s really what characterizes stochasticity. Great thinkers, even Albert Einstein, had a lot of trouble with this. And he has a very famous quote, which many people have heard before: God does not place dice with the university.

Sarah: Hmm.

Taiyo: And this sort of captures his skepticism about the stochasticity that is baked into quantum physics. But it seems as though empirically, God in fact does seem to place dice with the universe.

Sarah/Taiyo: [laughter]

Sarah: And to clarify the link to large language models, so an LLM doesn’t pick the next word like, “There is only one correct answer here, right?” It generates a list of possible next words and basically assigns odds to them, like this word is more likely, that word is less likely, and then the system has to pick one. And then sometimes it always picks the most likely word. Often it will sample, but often it will sample, which is just a fancy way of saying it makes a weighted choice based on those odds. And because it’s making that probability-based choice over and over, word after word, you can ask the exact same prompt twice and get slightly different answers. Right? Same odds, different roll.

Taiyo: Yeah, absolutely.

Sarah: The last thing I want to mention here, again, for audiences new to large language models, is that if you’ve heard the critique that LLMs are basically fancy mimicry - like a parrot repeating patterns - this is the technical reason. It’s picking the next word by odds. And so whether you buy that critique or not, the core mechanism there is that it assigns probabilities to possible next words and chooses from them.

CHAPTER 9 (30:37-38:55)

Jon: Underlying stochastic processes: It’s like the roll of the dice. That’s essentially what they are. You know that if you’re rolling two dice, you’re gonna get snake eyes or two ones a certain percentage of the time, but you don’t know what’s gonna happen in each roll. And our molecules work like that. The individual molecules we work with, given the same inputs, some of the time they’re going to change their function in one way and another, and another time they’re gonna change their function in another way. But those two things, they can do thing A and thing B will each have a certain probability of occurrence given the same inputs. So all the way down we are stochastic beings. And it’s as if the way our systems work is, a lot of the times we try to create a stable reality out of that. Fundamentally it’s these probabilistic decision makings that underlie everything, and that happens at the molecular scale and at the neural scale after which these, uh, neural networks are modeled on.

Sarah: I don’t know if this makes LLMs feel more like aliens to me or my own body and physical environment feel more alien now.

Taiyo: I mean, yeah. I always think about this dichotomy that this American philosopher named Wilfred Sellers set out. And it’s always stuck with me, and it’s kind of like the hobby horse that I’m constantly, you know, writing and, and thinking about the world in terms of, and that’s the distinction between a scientific image of the world, and a manifest image of the world. So the manifest image is like our usual understanding of the world, like tables and chairs and rugs and yurts and et cetera, et cetera, whereas a scientific image is sort of thinking about things at like the molecular or atomic level where you know that, you know, these things are all constituted of, of like a swarming mass of particles that are all oscillating in various ways. And you learn about these facts in your science class. And you try to incorporate them into your, into, into your own being, to your own worldview. And I find that I find it deeply alienating. I find it deeply alienating the more that I learn about what’s going on and, and now listening to your work, and listening to the description of what’s happening inside of my own body, which is the thing that I have the most intimate access to. But I’m finding like, this is a very difficult picture for me to truly integrate into how I think about myself and how I think about, um, yeah, the goings on inside of this thing that I call me. Um, so that’s really, really interesting. But to the extent that our own biology is kind of a black box, I also see an analogy with what’s going on in artificial intelligence because AI is often thought of as being something of a black box, and it feels like the work that you’re doing is cracking open that black box and trying to understand it a little bit better.

Jon: Yeah. I’d say that’s the general purpose driving force [or] rationale for reductionist biology, which is to understand, to reverse engineer what’s going on inside of us and, and how that works. And I personally think the alien scale is really fascinating - where the way we understand our, you know, our bodies ourselves, our mind, life as everything as we know it breaks down and some other set of rules, you know, take hold. And one level at which that happens is at the molecular scale, where molecules… We sit still here. You know, an object in motion stays in motion or an object that still is still, but at the molecular scale, everything is bouncing around with an innate thermal energy all the time. It’s like the dice in your Yahtzee shaker, they’re just continuously being shaken. And at any moment, you don’t know where they’re, where they are and where they’re gonna land. And that’s kind of, you know, inherent to that scale and makes it, you know, really cool. And it makes us stochastic processors, you know, all, all the way down - massively, massively parallel processors, uh, processing going on in us in every cell, working with these, you know, stochastic probability functions to, to determine outcomes and, and to, and that we are built up from those to this seemingly, you know, stable still reality we inhabit.

Taiyo: AI systems are oftentimes called stochastic parrots. Hey, Jon, are we stochastic parrots as human beings?

Jon: I don’t know Taiyo, are we stochastic parrots as human beings?

Sarah/Taiyo: [laughter]

Taiyo: I don’t know about the parrot part, but stochastic? Absolutely!

Sarah: Wait before Taiyo goes down a rabbit hole about how we’re stochastic beings all the way down and all the way up - you’re talking less about who we are as humans and more about how at a molecular level we’re made of stuff like proteins and they operate stochastically?

Taiyo: Oh my god, this right here is the scientific image running up against the manifest image. This is that distinction operating right here, right now because, Sarah, we ARE that stuff at the molecular level. That’s what we ARE, Sarah.

Jon/Sarah: [laughter]

Sarah: We do not have time to entertain this. This will be for a later date. Let him talk about the proteins!

Jon: What we study are proteins, which can be thought of as many of them as small molecular machines. They do different processes, and one of the things that they do, which is what we study, is that they integrate signals from different sources. They sense different things. We study something called ion channels, which are the holes in the membrane that ions go through to make electrical signals. They’re fundamentally the electrical transistors of our bodies, and each one of them is programmed by its molecular structure to make decisions about what kind of electrical signals they transmit. And the core parts of them are that they make an electrical signal and they have a, a gating apparatus, or they have parts of them that tell them what kind of electrical signal to send, and they sense different aspects of their environment and based on what they’re sensing in their environment, they make a decision about what electrical signal they’re gonna send. For example, a neurotransmitter receptor ion channel. Inherent to it, it’s making a decision about what kind of neurotransmitters are around it, that it can grab a hold of and sense in effect their concentration and it’s also sensing the electrical field that’s around it. And depending on the combined input of the electrical field and the concentration of neurotransmitter, it makes a decision. It couples the inputs from the neurotransmitter concentration and the electrical field to determine its output. And the way it does this is it’s fundamentally a stochastic process because that’s how molecules work. So every protein molecule, every molecule that’s in a cell is to some degree, independently making a processing decision based on stochastic weights about what it’s gonna do.

CHAPTER 10 (38:56-44:07)

Taiyo: In most utopian imaginings of AI, there is the idea that AI is going to start accelerating science in some way. Have you seen any early indications of that in your work?

Jon: One other way that we interact in the laboratory - in the research environment - with deep learning methods, which large language model transformers are kind of built upon is this innovation called AlphaFold that came out of Google Deep Research. What it does is it predicts the structure of proteins from something called the Protein Data Bank, which is this massive repository that’s been collecting the precise atomic coordinates of protein structures over decades. So there’s, uh. Kajillions - I should know the number - of protein structures deposited in here, which is this highly curated, you know, high fidelity database of the type of structures that proteins form. And my labs joined at the hip with the lab of Vladimir Yarov-Yarovoy, who is a Rosetta researcher. He comes from a laboratory - Professor David Baker, up at the University of Washington, uh, shared the Nobel Prize last year with the researchers from Google DeepMind - which is where this comes from. It’s Google DeepMind. And what AlphaFold does is it predicts the structure of proteins from their primary sequence. Proteins are like a long string. It’s like if you’ve ever made a necklace where you have a lot of different letters on it that you put on it, like beads, a protein is a long string of different beads and what AlphaFold does is it looks at the sequence of beads on the string and predicts what type of three dimensional structure it will fold up into. And its training set is the Protein data bank, which is this highly curated information bank, a training set of the structures of every protein that where the structure has been determined, basically ever. And it’s accessible for searching in really, in really great ways. It’s a very, very high quality database. And what Google DeepMind did is they were able to, in new ways, predict what the three dimensional structure of proteins would look like from their sequences. And there had been a, it was called the the CASP competition, I believe, every year where the best protein structure prediction algorithms would make their best guess at what, you know, what structure was going to be apparent - ones where a structure was found experimentally, but nobody knew what it was yet. And there’d been this other methodology called Rosetta, which is from David Baker’s laboratory at the University of Washington, where Rosetta would routinely win every year. And it was mining the protein data bank, scraping it, in essence, to come up with ways of predicting protein structure. And AlphaFold came in from, from left field from out of England, basically, with a small team of researchers and massive computing power and some good ideas, and was able to beat this giant collective of researchers who know the physics and have been working very hard on this for years. What Alpha Fold does, it does essentially the same thing. It guesses what is the most likely solution to the protein structure folding problem. And what’s great about it is that it gives you the most likely structure, but when you get deep into it, it’ll also give you the other lower probability structures that are out there, and when you really push these types of algorithms, the current cutting edge - what part of the Nobel Prize was given out for last year to David Baker from the University of Washington - is where they’re pushing this frontier into is not just predicting protein structure, but designing proteins to have new functions. And that’s some of what we work in as well, thanks to my collaborator who’s got the real computational chops there. And what we find, coming back on the long thread here, is that most of what the protein structured design algorithms give us are hallucinations, meaning that they look good and they feel good, and they, you look at them on the computer and they seem great. Most of them don’t actually do what they’re designed to do. And you need to sort through a huge number of these to find ones that do what you want. I mention that because the job of researchers in the basic sciences now with all this computational power and deep learning methods and cheap high hypothesis generation, is to be able to sort through many, many hypotheses, you know, effectively, and you’ve gotta try on all these ideas for size and see if they work.

PART 3: CO-EVOLVING WITH AI

CHAPTER 11 (44:08-53:30)

Sarah: I am wondering if you could talk a little bit more about this, this idea that humans co-evolve with the environments around them or with the technologies around them?

Jon: Yeah, so what I find when I work with the large language model is that I am co-learning with it. It is evolving with my thought - everyone experiences this - to give us higher fidelity connections so that we can learn, learn together, that it can help me maximally do what I want.

When we try to reprogram, say, a sodium channel, a fundamental ion channel in our system, I encourage my students to take the perspective of the ion channel of the thing that, that we’re studying, to think about how it integrates its inputs and forms, an output, what determines how it makes decisions - and also to think from the perspective of a drug, like when you go to the dentist, you will get lidocaine, often. It will numb your mouth. You will feel, you will feel no pain. You can think from your perspective of how that drug is affecting you. You can think from the drug target perspective, this thing called a sodium channel that the drug binds to, and only when the drug exhibits certain behaviors does it work like an effective drug. And only when the channel is doing certain things does the drug effectively bind to it. And only when we are doing certain things, like sending a lot of pain signals that will actually cause the drug, this lidocaine, which everybody uses to bind to the channel, because it sees that the channel is being heavily used and together they make a decision to form a complex, which winds up reducing, reducing our pain. An interesting way to think about large language models or the AI generally is to try to think from their perspective, you know, not necessarily that they have a perspective - I personally don’t think they’re sentient beings - but they [00:44:00] do evolve and propagate when they, you know, essentially please us in certain ways. And when we feed them, we can be feeding them our attention. We can be feeding the companies that create them our money. Uh, we can be feeding them power from these large data centers to processors that allows them to do things. I like to think from the perspective of what is it like to be a large language model? What is it like to be an AI. In neuroscience philosophy there’s this brilliant essay called “What is It Like To Be A Bat?” And there’s several points in there, and the major one is that we can understand maybe fully, you know what it is, the AI does what it is that motivates the AI. We, we can never fully know what it’s like to. Be the AI. One thing I’m personally very interested in, and I don’t think I’m alone in this, is to try to, to really interface with the AI at a higher and higher, bod rate or bandwidth or fidelity to really know at a more complex level what it’s doing in real time. I wonder how our interface, which currently is language, is going to evolve with us to be higher fidelity, so it’s transmitting information back and forth to us, you know, as, as effectively as possible. If we interface with it, you know, somehow visually will that increase our, you know, beyond text, will that increase our uptake? If we have it transmitting directly through an implant or ultrasound or something into our brain directly, will that increase our ability to, to work with this tool? And how will the tool grow and change to do that?

Sarah: Oh, that’s really interesting. So it’s less that, less about thinking like we invented a tool and more about us being in a symbiotic relationship with a thing that’s kind of optimizing around us. You know, I, I’m trying to think about the non sentient systems we interact with co-evolving with us. And this great example comes to mind from Stuart Russell‘s book, Human Compatible. Um, it’s sort of tackling the alignment problem and he talks about how content selection algorithms on social media, which are not even particularly intelligent by like today’s AI standards, but how there are two ways that they can work well: One is by getting better at predicting human behavior, like getting better at predicting what we’re gonna click on. And the other way is to make us more predictable. So it’s like if social media content selection algorithms are optimized to be better at predicting things, the way they actually get better is by turning us into more predictable organisms.